1290

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 24, NO. 3, AUGUST 2009

DC Power Flow Revisited

Brian Stott, Fellow, IEEE, Jorge Jardim, Senior Member, IEEE, and Ongun Alsa, Fellow, IEEE
AbstractLinear MW-only dc network power ow models are in widespread and even increasing use, particularly in congestion-constrained market applications. Many versions of these approximate models are possible. When their MW ows are reasonably correct (and this is by no means assured), they can often offer compelling advantages. Given their considerable importance in todays electric power industry, dc models merit closer scrutiny. This paper attempts such a re-examination. Index TermsCongestion revenue rights, contingency analysis, dc power ow, economic dispatch, nancial transmission rights, LMP pricing, unit commitment.

II. WHY DC MODELS? The linear, bilateral, non-complex, often state-independent, properties of a dc-type power ow model have considerable analytical and computational appeal. The use of such a model is limited to those MW-oriented applications where the effects of network voltage and VAr conditions are minimal (a very difcult-to-judge criterion). But then, as opposed to using the ac power ow model, the perceived advantages of a dc model are as follows. (a) Its solutions are non-iterative, reliable and unique. (b) Its methods and software are relatively simple. (c) Its models can be solved and optimized efciently, particularly in the demanding area of contingency analysis. (d) Its network data isminimal and relatively easy to obtain. (e) Its linearity ts the economic theory on which much of transmission-oriented market design is based. (f) Its approximated MW ows are reasonably accurate, at least for the heavily loaded branches that might constrain system operation. These are powerful attractions and, with exceptions to be noted are mostly valid. However, the big uncerlater, items and this complicates the choice betainty is proposition tween dc and ac models in any given application. On the other hand, sometimes there may be no viable alternative to the use of a dc model, for example when: (i) only linear theory and/or calculation techniques are available for certain (often market) applications; (ii) reliable voltage-VAr control data isnot available to support stable, meaningful ac power ow solutions; (iii) certain applications in large markets involve volumes of computing that would be prohibitive with ac modeling; (iv) a dc model is needed for cross-compatibility between two or more related applications. III. DC POWER FLOWA BRIEF BACKGROUND The term dc power (or load) ow comes from the old dc network analyzer [1], [2], in which each network branch was represented by a resistance proportional to its series reactance and each dc current was proportional to a MW ow. In the digital era this model becomes a simple, real (non-complex) nodal admittance matrix equation in terms of bus voltage angles and MW injections. The different dc model versions are distinguished by the definitions of the injections and admittances in this equation and, as will be shown here, minor variations in them can have big effects on model performance. Nevertheless, the original classical series-reactance version is still widely regarded as the dc power ow method. It is the version presented and derived in books dealing with power ow, for example [3][7]. Its adin the fast decoupled mittance matrixthe same as matrix

I. INTRODUCTION

HIS paper addresses so-called dc MW-only power ow modeling, which is of increased interest today because of recent upsurges in its usemostly in LMP-based market applications where prices are constrained by network congestion. Such applications include real-time security-constrained dispatch (SCED), day-ahead security-constrained unit commitment (SCUC), and the auctions and allocations associated with transmission rights (FTR-CRR-TCC). And more traditionally, dc models are widely used in contingency screening, transmission loading relief, transfer analysis, and medium-to-long term transmission planning. Many dc power ow model versions are available, but we have found nothing in the literature that identies and categorizes them. Papers that describe dc power ow applications frequently do not specify exactly which dc model was used. Dc power ow models are inherently approximate, and it is well known that their accuracies are very system and case dependent. At the same time, hard documentary data about this is sparse and often contradictoryfew large-scale dc model accuracy tests have been reported. Given the above, this paper offers two main contributions. Firstly, it reviews dc power ow methods by identifying and classifying different model versionsboth the presentation and some of the dc models are novel. Secondly, it summarizes the results of extensive, large-scale dc model testing, whose purpose was to investigate accuracy trends among the dc modeling variants. This paper covers dc modeling only at its fundamental level. It does not deal with other forms of linearization or the impact of any dc power ow model on any specic application.

Manuscript received December 10, 2008. First published May 19, 2009; current version published July 22, 2009. Paper no. TPWRS-00995-2008. J. Jardim and O. Alsa are with, and B. Stott is a consultant to, the Software and Information Systems Unit of Nexant, Inc., Chandler, AZ 85226 USA (e-mail: energysolutions@[Link]). Digital Object Identier 10.1109/TPWRS.2009.2021235

0885-8950/$25.00 2009 IEEE

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

STOTT et al.: DC POWER FLOW REVISITED

1291

Fig. 1. Series transmission line.

Fig. 2. Exact ac equivalent model of line.

power ow [8]yields MW sensitivities that have been applied very extensively by the industry [9][13], [21], [22]. Attempts at the theoretical error analysis of dc models have been made [14], [15]. However, it seems that a dc models performance on a non-small network can only be assessed experimentally, by comparing the MW ows obtained from dc and ac power ow solutions. A literature search reveals several reports of such experiments, all of them relatively recent [16][22]. Reference [16] compared dc power ow with other linear models on small networks. Reference [17] investigated ac-versus-dc LMP calculations on realistically large (13 000 bus) systems. Reference [18] and [19] describe systematic Monte Carlo simulations performed on small networks to investigate the inuences on the classical dc model of various ratios. Reference [20] studied LMP parameters such as line calculations with ac and dc models. Reference [21] studied the accuracy of power transfer distribution factors (PDTFs) on the European network. Reference [22] measured PDTF errors in small and large networks. In general, the above results tend to be more optimistic about dc model MW-ow accuracy than those from our own tests, to be described later. IV. DC POWER FLOW BASICS A. Notations Each network scalar is italic and lowercase. Thus, is a branch MW ow, is the angle in radians across the branch, and its inverse are series branch parameters on a 1 MVA per-unit base, and is a bus voltage magnitude in per-unit. The bus MW and angle vectors and are shown bold, italic and lowercase. The dc-approximation branch admittance is , and the corresponding dc nodal admittance matrix is . B. AC Branch-Flow Model The dc modeling process starts from the familiar ac transmission line shown in Fig. 1. The exact expressions for the sending and receiving end MW ows in the line are

C. Classical DC Model Derivation A very typical textbook dc power ow derivation [3][7] reduces the exact branch MW ow (1) via a sequence of approximations; thus

(2) (3) (4) (5) Equation (5) is the classical dc power ow model. As the basis for understanding and predicting dc model accuracy, this kind of derivation can be misleading. In contradiction to (2), it is essential to include an estimate of MW losses for a non-small power system. Then the approximations in (3)(5) above perform surprisingly better than might at face value be expected. This is explained briey in the Appendix. D. More General DC Modeling The dc modeling problem can be presented in a more general form as follows (this treatment is possibly novel). In the exact ac model of (1), let us designate

(6a) (6b) where the branch series loss is as . Then (1) can be rewritten

(7) (8a) (8b) and its equivalent circuit is shown in Fig. 2. Functions and would be sinusoidal if and remained constant. A dc branch model linearly approximates (7) and (8). Since this model must be bilateral, it is restricted to the form

(1a) (1b) For a transformer, each voltage magnitude is reected to the impedance side of its turns ratio. If the transformer has a . The terms involving phase shift , angle becomes conductance represent series branch losses. Branch shunt conductances are not shown heretheir inclusion is trivial.

(9) (10a) (10b)

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

1292

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 24, NO. 3, AUGUST 2009

the base point where it is constructed. Then any model errors will tend to increase only to the extent that the power system operating state moves away from this base point. F. DC Power Flow Matrix Equation Equations (9) and (10) lead to a dc nodal admittance equation with a sparse numerically symmetric matrix
Fig. 3. Dc equivalent model of line.

(12) where the components of the bus MW vector are generation; demand (load); bus shunt losses; hvdc infeeds; injections representing phase shifts; injections representing branch losses. V. DC POWER FLOW MODEL CATEGORIES A. Explicit Models These sparse non-incremental dc models are mostly used for the pre-contingency state in security-constrained applications. They can also be used for post-contingency power ow, although incremental versions are then more common. Here we introduce the terms hot and cold start. a) Hot-Start Models: This type of model is constructed at a solved ac power ow base point. It is often used in real-time security constrained economic dispatch (SCED), run from a state estimator solution [23][25]. It can also be used in short and medium term operation and planning studies where an initial ac solution is available. b) Cold-Start Models: A dc model has to be constructed in cold-start mode when a reliable base-point ac power ow solution is unavailable (usually due to lack of good voltage/VAr data). This situation typically arises in dc-model-based security constrained unit commitment (SCUC) [26][28], FTR-CRR auctions and allocations [29][31], and longer-term planning studies [32], [33]. B. Incremental Models These models compute changes from a known ac or dc basepoint state. By denition they t the initial base point. There are two types. a) Sparse Models: These models are used for real-time SCED and any other application where a solved base-point ac or dc model is available. b) Sensitivity Factor Models: These factors are generated from the sparse dc network matrix. When used directly in system contingency monitoring and remedial control action, they are pre-computed. Their names vary. VI. HOT-START DC MODELS An initial solved ac power ow solution is very helpful as a base point from which to construct an explicit dc model. This section describes several of the many possible model variants. Series and shunt MW losses and ZIP loads are evaluated from the base point ac solution and generally remain xed thereafter

Fig. 4. Graphic of (1a) showing dc model.

In (9) is the xed dc equivalent branch admittance. In (10) and are xed injections whose sum approximates the branch loss at some chosen operating point. This translates to the dc equivalent circuit of Fig. 3. E. DC Modeling Challenge Fig. 4 depicts a section of the true nonlinear curve versus . The intercept on the axis is , which for small remains close to . The intercept on the -axis is . That is, the curves asymmetric roughly displacement from the origin is strongly a function of both the branchs impedance and its state-varying voltage drop. The versus is very similar. curve The dashed straight line in Fig. 4 represents the dc model approximation to the true curve of versus . Ideally, we would and for which (9) and (10) prolike to nd values of vide the best linear t to the true curves over the expected range of practical operation. Unfortunately, identifying a good such model is complicated and ends up being largely empirical. Candidate denitions for admittance are suggested directly by (3)(5) (assuming certain xed voltages in the rst of and , these); or could be chosen as a value in between ; or we could obtain from the slope of the such as tangent to the true curve at a suitable point (11) Likewise, by comparing (10) with (8), the most obvious choices for loss-approximating injections and would appear to be and , evaluated at a suitable point. Intuitively, it seems desirable, particularly in real-time/online applications, for the dc model to be as accurate as possible at

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

STOTT et al.: DC POWER FLOW REVISITED

1293

in (12). At the base point, all hot-start models have the same total MW losses as the ac solution, and therefore their reference bus injections are correct. A. Net Loss Dispersal A dispersed-loss dc model ignores (10) and their constants and . It can be constructed at an ac base point by semi-arbitrarily distributing the known net losses as additional bus injections at the load and/or generation buses to form vector in (12). After solving for the bus angles, (9) provides the branch ows. Such a hot-start model has little attraction, however, because the exact amounts and locations of the initial losses are known and usable, as follows. B. Localized Loss Modeling Branch constants and in (10) can be dened from the known individual losses in different ways, for instance: 1) as and , respectively, evaluated from (6) at the initial ac solution point (by comparing Figs. 2 and 3); 2) as 50% each of the initial branch loss. The s form vector . After solving (12) for the bus angles, the branch ows are given by (9) and (10). C. Base-Point Matching When the dashed line in Fig. 4 passes through the point, the localized branch loss model matches the MW ows and angles of the ac base point perfectly. Thus (this seems to be unpublished) we will equate (10a) with (8a) at this point thus:

A. Net Loss Dispersal This commonly-used approach is similar to that in Section VI-A, except that here the net loss has to be estimated, for instance as a percentage of net load. In routine operational simulations on a given system, net losses can be guessed fairly well. But in any case, the loss dispersal is so arbitrary that these loss estimates have second-order effects on MW ow accuracies (sometimes even a zero estimate is used). The net estimated loss is distributed as injections at the load in (12). Island net and/or generation buses to form vector MWs are then balanced by scaling or otherwise adjusting the generations and/or the loads. (Caveat: if most units start at maximum, any net generation increase could end up unevenly allocated to just a few units.) The choice of dc branch admittance is usually limited to a state-independent value such as or . ZIP loads may be evaluated at nominal or typical voltages. Shunts are evaluated likewise or are simply lumped with the net loss. Then (12) is solved for the dc-model base-point angles, and (9) gives the branch ows. B. Loss Redistribution Once constructed and solved, the dispersed-loss dc model of Section VII-A can be rened by redistributing the losses to the individual branches as follows. (a) Calculate loss components and for each branch, eior by setting both as [4]. ther from (6) with (b) Scale each in order to keep the net system loss at its original estimated value. in (10) as and , respectively, (c) Set loss terms and establishing a new vector . (d) Solve (12) and obtain new branch ows from (9). The above process can be iterated, and convergence is rapid. A variant is to omit step 2), allowing the net branch losses to change, in which case the system MWs have to be rebalanced at each iteration. This variant [20] usually converges but the net system loss has no anchor value and is free to drift. C. Fixed-Voltage AC Solution Start Another approach to cold-start modeling is to obtain the solution of a simplied ac power ow problem, to which a matched hot-start dc model can be tted. All buses in this simplied ac formulation have their voltage magnitudes xed at nominal or typical values. That is, all buses become designated as PV with no VAr limits. The basic at-voltage version, requiring no voltage/VAr data whatever, sets all voltages and taps to 1 per-unit. This ac solutions VAr ows are completely wrong, of course. However, there is some hope that its MW ows, net losses and loss distribution will be better than those in Section VII-A. VIII. INCREMENTAL DC MODELS A. Sparse Matrix Models The incremental version of (9) is (15)

(13a) and we equate (10b) with (8b) as thus:

(13b) Two ways of achieving this matching are as follows. a) -Matching: With any specied , (13) can be solved and . Note that their sum to provide matching values for remains equal to the initial branch loss . This is also equivalent to adding a xed matching phase shift to each network branch. and either (13a) or b) -Matching: For any given (13b) can be solved for the matching value for as follows:

(14) as . This model was used (by one where of the present authors) in the software of [34]. VII. COLD-START DC MODELS The absence of an initial ac power ow solution makes constructing a reliable base-point dc model much more difcult than in the hot-start case.

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

1294

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 24, NO. 3, AUGUST 2009

where superscript signies the base-point value. In matrix form, this becomes the incremental version of (12)

IX. SOME OTHER DC MODEL ISSUES A. Phase Shifting Transformers The dc equivalent model for a phase shifting transformer with xed angle and admittance is trivial. It comprises a pair of at the branch terminals. Variation of MW injections with angle is usually not modeled. The big modeling problem occurs for a phase angle regulator (PAR), where is automatically adjusted to maintain a scheduled MW ow . When is in range, the simplest (not necessarily best) way of modeling this is as an open circuit with . However, when the angle reaches terminal injections of a limit, the device reverts to its xed-angle version. The network model and its sensitivities therefore change each time an angle is xed on or backs off a limit, and the now-discontinuous linear dc model has to be solved iteratively. Iteration can only be avoided by ignoring phase angle limits, but this is likely to produce grossly unrealistic MW ows. In an optimizing calculation, one attempt to circumvent the problem is to designate each phase shift angle as an optimization variable, and constrain the MW ow. This brings its own complications, in terms of costs of shift, target ranges, binding constraints and solution uniqueness. Thus, the modeling of PARs with angle limits has huge analytical and computational consequences. It makes pre-calculated sensitivities virtually unusable, at least for the pre-contingency network state (post-contingency phase shifts are often represented as xed at their pre-contingency values). Similar considerations apply to HVDC and FACTs devices that automatically control MWs. B. Security-Constrained Optimization The biggest use of dc-type models is in linear SCOPF/ SCEDsecurity-constrained optimal power ow or dispatch [23][25]. Similar models are used in security constrained unit commitment [26][28] and nancial transmission rights applications [29][31]. This is not the subject of the present paper, but a few comments on the dc modeling aspects are appropriate. Formulation: A dc-based SCOPF formulation can be expressed in its simplest form as the minimization of (19) subject to linear sets of constraints, for (20a) (20b) where superscript signies the pre-contingency state and otherwise refers to contingency case i. Discontinuities: As described in Section IX-A, iteration is required whenever the continuity of the linear dc model is interrupted. This sacrices some of the advantages of linear modeling and makes solutions more complicated, time-consuming, path-dependent and non-unique. Apart from PAR angle limiting, other model discontinuity examples are when 1) a MW

(16) and By substituting we get the equivalent explicit dc model into this equation,

(17) . which is identical to the -matched model of Section VI-C For calculations (e.g., contingency analysis) involving changes in the network, (16) can be written

(18) Branch and bus outages (single or multiple) are handled efciently by compensation or factor-updating techniques [35]. These techniques can also handle more complex topology changes; otherwise matrix re-factorization is undertaken.

B. Sensitivity Factor Models Sensitivity factors are very widely known and used [5], [9][13], [21], [22]. Their pure dc versions derive directly from (16) and (18). The factors of main interest are usually known as: the MW PTDF (power transfer distribution factor) change in a branch ow for a 1 MW exchange between a bus and the point or distributed reference; the post OTDF (outage transfer distribution factor) contingency MW change in a branch for a 1 MW pre-contingency bus-to-reference exchange; the MW change LODF (line outage distribution factor) in a branch ow due to the outage of a branch with 1 MW pre-outage ow. The PTDF and OTDF factors have been used explicitly by system operators for transmission loading relief (TLR) procedures. Most network security-constrained optimizations internally calculate and use these factors. The LODFs can provide very fast dc contingency screening. Once computed, it is necessary to store and use only the nonsmall factors, which are relatively very few in number. That is, only the transmission elements that are sufciently sensitive to a given contingency are monitored. As long as the network topology does not change, the factors remain the same and they can be used repeatedly and rapidly. But whenever the network changes, they have to be recalculated. OTDF and LODF factors can readily be derived and applied for contingencies comprising multiple line outages. However, the efciency of the LODF approach is lost when any contingency involves changes other than simple non-islanding series reactance outagesfor instance the outage of a phase shifter, or any outage requiring bus MW redistribution.

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

STOTT et al.: DC POWER FLOW REVISITED

1295

outage or islanding requires MW rebalancing within generator limits, 2) secondary switching is modeled, or 3) market-specic precedence and/or infeasibility rules are invoked. Normally, a dc-modeled SCOPF formulation would have to be simplied to the point of extreme impracticality before it could be fed into a general-purpose optimization package. Consistency: Given the vast numbers of linear network constraints in (20), the SCOPF/SCED solution process needs to monitor them in an outer-loop, which feeds any violated or near-violated constraints to the redispatch engine. Quick-anddirty implementations perform this outer loop once only, ignoring the fact that redispatch to address some but not all insecurities may create even worse insecurities. Any trustworthy SCOPF must iterate outer-loop monitoring until no insecurities remain. Such iteration (generally taking just a few passes) requires consistent pre- and post-contingency dc models. Otherwise, constraints enforced in one pass will not be on their limits in the next pass, and oscillatory solutions can result. X. ASSESSING DC MODEL ACCURACY A. General It is axiomatic that dc power ow approximations vary enormously in accuracy for different systems, loadings, ow patterns and individual transmission elements. In this volatile, analytically confusing, heuristic area, there is no prospect for simply ranking dc models in order of goodness. Nevertheless, by running thousands of cases on dozens of networks, we have tried to accumulate some general insights into the expected accuracies of different dc models. Most dc models nd their use within some form of security-constrained optimization. Therefore it seemed appropriate to focus on a models accuracy: at the base point where the model is constructed; after contingencies; after redispatch. Fig. 5 outlines our test procedure. In this eld, small differences in methodology can lead to big differences in results. Therefore we try to explain our test procedures in some detail. Every pre- and post-contingency power ow solution was run with both ac and dc models, making sure that the comparisons are realistic. In particular, the slack power in every ac solution was shared among all generators, so that MW-loss changes do not accumulate at the reference bus. B. MW-Flow Comparisons The scope of our tests was very basicto measure the errors in the dc-modeled MW ows. We did not address the techniques for imposing limits (mostly quoted in MVA or amperes) on these ows. Typical such methods try to account for the MVAr ows that are absent from dc modelsfor example, a power factor is imputed to each branch ow. Reference [36] describes a more sophisticated approach. C. MW-Flow Accuracy Measurement The critical ows in a network are the potentially congestive onesthose that can substantially affect system dispatch and

Fig. 5. Test procedures.

TABLE I POWER FLOW TEST SYSTEMS

pricing. Thus, in all our tests, we monitored branches loaded above 70% of rating (and we ignored all ows below 50 MW). The error in each monitored dc ow was expressed as a percentage of the corresponding exact ac MW power ow. At no stage in the tests were we able to discern any statistical patterns in the dc-ow error scatters. This defeated all our attempts to nd concise, meaningful indices with which to characterize and display dc-model accuracies. We ended up simply tabulating the ranges (i.e., the extremes) and the averages of the dc-model ow errors. We separately recorded the average positive errors (dc MWs too high) and the average negative (dc MWs too low) errorsunder-estimating critical ows compromises system security, while over-estimating them leads to transmission under-utilization and congestion over-pricing. D. Test Systems Tests were carried out on dozens of power systems. Table I lists the six systems for which illustrative results are given here. The voltage ranges dened as Low, High, and Extra-High (LV, HV, EHV) are of course somewhat arbitrary. H, M, and L, respectively, stand for Heavy, Medium, and Light load. Each such system is a large modern power grid, reecting the tendency towards centralized modeling and large markets. It also exposes dc modeling to a diverse range of network characteristics. The data for each system comes as a solved ac power ow case that has been widely used by the industry, either in its entirety (as tested here) or in reduced form. We ltered out of

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

1296

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 24, NO. 3, AUGUST 2009

Fig. 7. Detailed error spread corresponding to System 1 in Fig. 6. Fig. 6. Base-point cold-start dispersed-loss model.

our error analyses a few data items that were obviously wrong, such as base-point branch ows in Systems 1 and 2 that were 100% overloaded. XI. TEST-BASED OBSERVATIONS This section offers comments on the dc-modeling trends that we have observed in our tests, with illustrative results. A. Base-Point Modeling As previously mentioned, the base-point accuracy of a dc model is importantlarge errors at the models base point will rarely diminish much when the operating point changes. Loss Dispersal Models (Sections VI-A and VII-A): The most common explicit dc models are those that distribute net losses in some semi-arbitrary manner among the network buses (or ignore losses completely). By their nature, these models are not accurate in large power systems. Fig. 6 shows the dc MW errors for a cold-start model with (generally a bit better than or ). In this specic case, net losses, underestimated by 10% relative to the exact ac reference solution, were dispersed to the load buses. In Fig. 6, for the bar corresponding to a given test system and kV range, the gray portion shows the maximum MW errors. The black portion depicts the average of such errors. These results are little different from those of the corresponding dispersed-loss hot-start model, constructed with accurate net losses, etc. They are not much worse when the losses are estimated as zero. The average errors in Fig. 6 seem reasonable, but the extreme ones in several of the test systems are certainly not. The large errors are not just isolated casesthey are found throughout the network at the different voltage levels. This is illustrated in Fig. 7, which plots the full set of monitored ow errors in the results of Fig. 6 from System 1. On systems with long-distance generator-to-load ows, the dispersal of losses to the generator buses worked slightly better than dispersal to load buses, but not by much. Loss dispersal to all buses likewise had minor effects. Localized-Loss Hot-Start Models (Sections VI-B and C): Compared with the dispersed-loss models of Section VI-A, the localized-loss versions of Section VI-B produced huge improvements (e.g., halved errors). In the latter, incorporating

the initial voltages from (3) into the dc admittance often halved the MW ow errors again. A small further improvement was to dene as the tangent to the initial curve per (11). The denitions in Section VI-B performed amazingly better than those in Section VI-B . This is useful dc modeling insight, but from a hot-start perspective, it seems always preferable to use the matched versions of Section VI-C, which t the ac base point MW-ow and angle solution exactly. Depending on the implementation, this is also true of the incremental models of Section VIII. Cold-Start Model Enhancements (Sections VII-B and C): We found that the loss-redistributing steps 1)4) of Section VII-B improved the cold-start dc model signicantly. A single iteration was sufcient. Disappointingly, System 1 and 5 each had one persistent large MW-ow error, attributed to ow-path -ratio differences. Extremely similar results were obtained using the at-voltage ac technique of Section VII-C, including the identical large anomalous errors. The associated dc model average and worst errors are shown in Fig. 8, which is directly comparable with Fig. 6. (Note: in a xed-voltage ac formulation, ows on certain branches with impractically high ratios are infeasible. Temporarily reducing these ratios to 1 overcame any convergence problems.) Omitting step 2) in Section VII-B lets the losses nd their own level and sometimes gives small further improvements. But in other cases a big positive feedback effect takes place, and with iteration the errors increase enormously on several highloss branches. Extra logic might take care of this. Sources of Large Errors: In all results, considerable efforts were made to pinpoint the sources of the largest dc MW ow errors. By far the most common situation is when heavy ows ratios differ divide themselves between local paths whose widely. Branches with very large transmission angles can be another problem area. Signicantly, we also often saw large errors on branches with no bad characteristics, indicating propagation from the poor linear modeling of other branch ows. B. Post-Contingency Accuracies Contingencies Simulated: On each test system we outaged every branch whose base-point loading exceeded 70%. In the ac solutions, post-contingency controls on taps, phase shifts and switched shunts were blocked. Any non-convergent outage

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

STOTT et al.: DC POWER FLOW REVISITED

1297

Fig. 8. Base-point at-voltage ac start.

Fig. 10. Cold-start contingency error ranges.

Fig. 9. Hot-start contingency error ranges.

case was necessarily ignored. Likewise, those few outages producing implausible ac voltage drops or overloads were ignored, on the basis that they would most likely be accompanied by supplementary control actions. When in the dc solutions the losses of outaged branches were also outaged, they had to be distributedotherwise they were occasionally large enough to distort the ows near the reference bus. Hot-Start Contingency Modeling: Modeling errors that are contingency-specic can most easily be seen by starting from a hot-start dc model that perfectly matches the ac ows. The following results are based on the -matching version of Section VI-C. The -matched version gave very similar results. Similar or identical results are obtained from the incremental versions of Section VIII. Fig. 9 gives results using the denition for in (11), which generally performs best. However, its state-dependent seems less suitable for LODF versions, whose factors are intended for multiple reuse. The gray area of each bar in the gure represents the worst errors over all contingency cases. The black portion of the bar delineates the average of the worst errors among all contingency cases. These averages are very small. They are dwarfed by error outliers that seem to be due to ow division between parallel paths with very different nonlinearities, prinratios. cipally different Another lesser cause of post-contingency errors was due to the dc models xed-loss modeling. Occasionally, local pre-contingency branch losses are high, and after the contingency they become very low, or vice versa.

Fig. 11. Hot-start re-dispatch.

Cold-Start Contingency Modeling: With a standard cold-start model as per Section VII-A, the average-worst and globally-worst post-contingency errors tend to be very high. This is illustrated in Fig. 10, which very non-rigorously may be thought of as a combination of the results in Figs. 6 and 9. These errors can be partially mitigated by starting with a at-voltage ac solution, for instance. C. Post-Redispatch Accuracies Our interest here was in observing the changes in accuracy of the base-point dc models after non-traumatic operating changes such as routine constrained re-dispatch during load following. There are unlimited alternatives for performing such simulations. In our case, we simply ran each system with its load level reduced by 5%, randomly re-dispatching each generator within a bound of 25% of its output. Fig. 11 gives error results for the hot-start base-point model with -matching and as tangent per (11). It is seen that for such non-traumatic operating changes, the errors are relatively small. Obviously, different random re-dispatching produces slightly different numbers. Of the hot-start models, this version tends to be the most accurate one. Incremental model results are quite similar. When the same load-reduction/re-dispatch exercise is conducted on a cold-start dc model, the models base-point error

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

1298

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 24, NO. 3, AUGUST 2009

lapse. Then the difference between the ac and dc ows can be considerable. In contingency analysis, this was deemed most likely to be a result of nave ac modeling. Paradoxically, the dc results in these extreme situations sometimes seemed more realistic than the ac results. This is a very gray area, and it underscores the need for detailed analysis of dc modeling in any specic power system and network application. Extremely occasionally, cases of system near-separation arise, and the dc-model angle across a branch becomes impractically large. These pathological cases are detectable. XII. GENERAL COMMENTS
Fig. 12. Hot-start re-dispatch

contingency analysis.

Fig. 13. Cold-start re-dispatch

contingency analysis.

results (not shown here) dominate, looking much like Fig. 6. However, they were fractionally smaller, attributable to reduced nonlinearity because of lower load. The results for contingency analysis at the hot-start re-dispatched points of Fig. 11 are shown in Fig. 12. The corresponding results for contingency analysis following re-dispatch of a cold-start model are shown in Fig. 13. This is very similar to Fig. 10, and once again illustrates the problem of constructing viable cold-start models. D. Reliability Ac power ow fails to converge in sufciently severe contingency and re-dispatch cases. On the other hand, dc power ow inherently always solves. Such a dc property can clearly be very dangerous. However, it may actually be valuable in some network-constrained applications. Consider security-constrained dispatch with an ac network model. Whenever the power ow solution fails (pre- or postcontingency) the relevant security limits cannot be monitored. And network constraints cannot easily then be imposed on the dispatch process in order to avoid operation in a region that is unstable algorithmically, analytically and/or in the real-life system. A dc network model does not have this problemit always provides the dispatch process with meaningful (if not necessarily accurate) MW-related constraints. Among converged ac solutions we encountered occasional situations where a bus voltage (generally at LV) was near col-

As shown here, a spectrum of dc modeling variants is possible, with accuracies that vary enormously over different networks, transmission elements, and loading levels. Inevitably, however, certain dc model versions are superior to others in given circumstances, and trends can be observed. As previously noted, dc-model accuracy is of great interest because network constraints that erroneously become congestive, or fail to do so, might have huge impacts on market costs, as well as on system security monitoring. When assessing the accuracy of a dc power ow model against an ac model, it must be recognized that the latter has statistical and possibly gross errors. This does not affect the basic assessment. Our test-based observations indicate that critical (potentially binding) dc MW ows are on average wrong by plus or minus a few percent, with little overall bias towards over- or under-estimation. It is very clear from our results, however, that the worst errors on such critical ows can sometimes be alarmingly high. (We have carefully veried the correctness of our calculations with different totally independent software packages.) How often, and by how much, such error outliers will distort system dispatch and pricing is a subject for analysis outside the present papers scope. Clearly, there is a big accuracy gap between hot-start and cold-start models. In the former, matched and incremental dc models work best, particularly in moderate-change redispatch and contingency analysis. On the other hand, cold-start models are seen, even from the few results shown, to be vulnerable to huge inaccuracies on certain critically-loaded branches. This seems not to be widely recognized [28]. It seems important to identify the trouble spots in a network where dc-model error extremes can appear. It might then be possible to correct data anomalies, use better dc models or apply special mitigations. Sometimes it is expedient to omit lower voltages from the network model, or at least not to monitor offending but unimportant circuits. Best-practice ac power ow modeling at all voltage levels is imperative to minimize such trouble spots. We ourselves noted and xed a number of evident data problems in various power ow models. Certain types of network reduction produce highly unrealistic equivalents, and should be avoided. Obviously, the big enemy of dc modeling is nonlinear network MW behavior. An excellent illustration of the nonlinearity associated with network stress is the difference in accuracies between Test Systems 5 and 6, which represent the same system under heavy and light loading, respectively.

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

STOTT et al.: DC POWER FLOW REVISITED

1299

Compared with classic textbook dc power ow, which looks almost trivial to code, many dc versions require careful implementation. For instance, a branchs sending and receiving dc ows may be unequal, and balancing an islands MWs (including slack redistribution) is often needed. Locally-controlled phase shifter, HVDC, and FACTs devices considerably complicate the use of dc models, because recursive solutions are needed to impose or back off their limits. The same applies to any other discontinuities, such as post-contingency MW redistribution to generator units within their limits. Adding to this the fact that contingency constraints are so numerous that they must be handled in outer calculation loops, there seems to be little scope for solving realistic dc-modeled problems at one pass with a general-purpose linearly-constrained solver. XIII. CONCLUSION The electric power industry has invested heavily in the use of dc-type power ow models for network security and pricing calculations. However, our results underscore that the accuracy of a particular dc model, or indeed of dc modeling in general, should never be taken for granted in any given power system and application. Dc-model testing of the kind described here is not very complicated. When it reveals large MW ow errors on potentially binding network elements, there is a strong case for investigating the source of such errors and how they impact LMP prices and/or security as appropriate (this requires analytical software that supports both ac and dc network models). As long as certain power system applications continue to rely on linear network models, dc-type modeling will remain of high interest. We hope that the present paper might stimulate further development, testing, and verication in this area.

is 3) When substituting (5) for (4), the percentage error in . Thus, for and 3, respectively, the errors are 11%, 100%, and 900%. In practice, the errors in the dc ows tend to be much lower than would be predicted from 1)3) above. The reason is that the dc model is essentially a direct-current dividing network, allowing MWs to ow under Ohms and Kirchhoffs laws according to the relative values of the dc branch admittances . These admittances only need to be in approximately correct proportion to each other. Thus, if all voltages are 1.5 per-unit, the actual MW ow error associated with 2) above is zero. If the line is radial, no MW ow error apart from losses will arise as a result of or . The same is true for any radial chain of branches (in the absence of ZIP loads). By the same token, when dc MW ows divide themselves among network paths whose true ac models are substantially different in nonlinearity, large errors can result. Moreover, these errors can be unpredictably self-cancelling or cumulative. In the latter case, they can propagate round network loops and degrade the MW ow accuracies on branches whose individual dc models are excellent. REFERENCES [1] W. C. Hahn, Load studies on the D-C calculating table, General Electric Review. Schenectady, NY, vol. 34, pp. 44445, 1931. [2] J. A. Casazza and W. S. Ku, The coordinated use of A-C and D-C network analyzers, in Proc. Amer. Power Conf. Illinois Institute of Technology, Chicago, IL, 1954, vol. 16, pp. 44753. [3] U. G. Knight, Power Systems Engineering and Mathematics. New York: Pergamon, 1972. [4] A. J. Monticelli, Load ow in electric energy networks, Editora Edgard Blcher Ltda (in Portuguese) 1983, pp. 2527. [5] A. J. Wood and B. F. Wollenberg, Power Generation, Operation and Control. New York: Wiley, 1996. [6] L. Powell, Power System Load Flow Analysis. New York: McGrawHill Professional Series, 2004, ch. 11: DC Load Flow. [7] A. Gmez-Expsito, A. J. Conejo, and C. A. Caizares, Electric Energy Systems Analysis and Operation. Boca Raton, FL: CRC, 2009. [8] B. Stott, Review of load ow calculation methods, Proc. IEEE, vol. 62, pp. 916929, Jul. 1974. [9] P. Sauer, On the formulation of power distribution factors for linear load ow methods, IEEE Trans. Power App. Syst., vol. PAS-100, pp. 764779, Feb. 1981. [10] NERC Minutes, Distribution task force meeting, Mar. 2324, 1999. [Online]. Available: [Link] [11] S. Greene, I. Dobson, and F. L. Alvarado, Sensitivity of transfer capability margins with a fast formula, IEEE Trans. Power Syst., vol. 17, no. 1, pp. 3440, Feb. 2002. [12] X. Cheng and T. J. Overbye, PTDF-based power system equivalents, IEEE Trans. Power Syst., vol. 20, no. 4, pp. 18681876, Nov. 2005. [13] R. Baldick, T. Overbye, and K. Dixit, Empirical analysis of the variation of distribution factors with loading, in Proc. IEEE PES General Meeting, 2005, pp. 221229. [14] R. J. Kane and F. F. Wu, Flow approximations for steady-state security assessment, IEEE Trans. Circuits Syst., vol. CAS-31, no. 7, pp. 623636, Jul. 1984. [15] R. Baldick, Variation of distribution factors with loading, IEEE Trans. Power Syst.., vol. 18, no. 4, pp. 13161323, Nov. 2003. [16] P. Yan and A. Sekar, Study of linear models in steady state load ow analysis of power systems, in Proc. IEEE PES Winter Meeting, 2002. [17] T. J. Overbye, X. Cheng, and Y. Sun, A comparison of the AC and DC power ow models for LMP calculations, in Proc. 37th Hawaii Int. Conf. System Sciences, 2004. [18] K. Purchala, L. Meeus, D. Van Dommeln, and R. Belmans, Usefulness of DC power ow for active power ow analysis, in Proc. IEEE PES General Meeting, Jun. 2005. [19] D. Van Hertem, J. Verboomen, R. Belams, and W. L. Kling, Usefulness of DC power ow for active power ow analysis with ow controlling devices, in Proc. 8th Int. Conf. AC and DC Power Transmission, Mar. 2006.

APPENDIX COMMENTS ON DC MODELING APPROXIMATIONS The approximations used in deriving (2) to (5) provide certain insights but also some false impressions about the dc modeling process. Firstly, neglecting the losses as per (2) on a branch implies a typical error of only a few percent. However, in a non-incremental dc mode, such errors are cumulativethey show up at the reference bus. For a non-small power system the MW ows on the branches in the buss vicinity can be wrong by hundreds of percent. The dc model must then preserve the system generation-load-loss MW balance by representing losses as equivalent injections, either dispersed among the buses or local to each branch. Secondly, at rst sight it would seem that the approximations in (3)(5) can sometimes introduce enormous errors into the dc model. To illustrate: and is 8.6%. 1) At 40 the difference in (3) between (Angles as high as this may be found on very long lines.) per-unit, the approximation producing 2) If (4) has a 44% error. (Typical NERC-MMWG models have voltages in the range 0.75 to 1.4 per-unit.)

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.

1300

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 24, NO. 3, AUGUST 2009

[20] F. Li and R. Bo, DCOPF-based LMP simulation: Algorithm, comparison with ACOPF, and sensitivity, IEEE Trans. Power Syst., vol. 22, no. 4, pp. 14751485, Nov. 2007. [21] C. Duthaler, M. Emery, G. Andersson, and M. Kurdizem, Analysis of the use of PTDF in the UCTE transmission grid, in Proc. 16th PSCC, Glasgow, U.K., Jul. 2008. [22] M. Liu and G. Gross, Effectiveness of the distribution factor approximations used in congestion modeling, in Proc. 14th PSCC, Seville, Spain, Jun. 2002. [23] B. F. Wollenberg and W. O. Stadlin, A real-time optimizer for security dispatch, IEEE Trans. Power App. Syst., vol. PAS-93, pp. 16401649, 1974. [24] B. Stott, J. L. Marinho, and O. Alsa, Review of linear programming applied to power system rescheduling, in Proc. IEEE PICA Conf., Cleveland, OH, May 1979. [25] Economic Dispatch of Electric Generation Capacity, Report to Congress, Feb. 2007, U.S. Department of Energy. [26] F. N. Lee, J. Huang, and R. Adapa, Multi-area commitment via sequential method and a dc power ow network model, IEEE Trans. Power Syst., vol. 9, no. 1, pp. 279287, Feb. 1994. [27] M. Shahidehpour, H. Yamin, and Z. Li, Market Operations in Electric Power Systems. New York: Wiley, 2002. [28] H. Pinto, F. Magnago, S. Brignone, O. Alsa, and B. Stott, Security constrained unit commitment: Network modeling and solution issues, in IEEE PSCE Conf., Oct. 29Nov. 1, 2006, pp. 17591766. [29] W. Hogan, Financial Transmission Right Formulations, Tech. Rep., 2005. [Online]. Available: [Link] [30] V. Sarkar and S. A. Kharparde, A comprehensive assessment of the evolution of nancial transmission rights, IEEE Trans. Power Syst., vol. 23, no. 4, pp. 17831795, Nov. 2008. [31] O. Alsa, J. Bright, S. Brignone, M. Prais, C. Silva, B. Stott, and N. Vempati, FTRsThe rights to hedge congestion costs, IEEE Power Energy Mag., vol. 2, no. 4, pp. 4757, Jul./Aug. 2004. [32] X. Wang and J. R. McDonald, Modern Power System Planning. New York: McGraw-Hill International Editions, 1994. [33] A. Santos Jr., P. M. Frana, and A. Said, An optimization model for long-range transmission expansion planning, IEEE Trans. Power Syst., vol. 4, no. 1, pp. 94101, Feb. 1989.

[34] Development of Production Grade Software for the Evaluation of Simultaneous Transfer Capability (TRACE), EPRI, Project WO3140-06, 1992-5. [35] F. L. Alvarado, W. F. Tinney, and M. K. Enns, Sparsity in largescale network computation, in Book: Advances in Electric Power and Energy Conversion System Dynamics and Control. New York: Academic, 1991/3. [36] S. Grijalva, P. W. Sauer, and J. D. Weber, Enhancement of linear ATC calculations by the incorporation of reactive power ows, IEEE. Trans. Power Syst., vol. 18, no. 2, pp. 619624, May 2003. Brian Stott (F83) received the Ph.D. degree from the University of Manchester, Manchester, U.K. He taught in several universities and worked at CEPEL in Brazil. In 1984, he co-founded PCA Corporation in Arizona and is currently a Principal Consultant to Nexant, Inc., Chandler, AZ. Dr. Stott is a Member of the U.S. National Academy of Engineers and an IEEE Millennium Medalist, and he participates in the IEEE Distinguished Lecturer Program.

Jorge Jardim (SM93) received the Ph.D. degree from Imperial College, London, U.K. He worked in Brazil (FURNAS, CEPEL, and ONSthe National System Operator), Canada (BC Hydro), and now in the U.S. He developed the static and dynamic security software used online in BC and Brazil. He is a principal engineer and project manager at Nexant, Inc., Chandler, AZ.

Ongun Alsa (F95) received the Ph.D. degree from the University of Manchester, Manchester, U.K. He worked in Turkey (METU), Brazil (CEPEL), and the U.S. In Arizona in 1984, he was a co-founder of PCA Corporation. He is currently Senior Vice President and Director of the Software and Information Systems Unit of Nexant, Inc., Chandler, AZ, responsible for all software R&D and products.

Authorized licensed use limited to: DELHI TECHNICAL UNIV. Downloaded on August 26, 2009 at 10:25 from IEEE Xplore. Restrictions apply.