Power Predictor Engine: System Architecture & Reference

Overview

The Power Predictor Engine is an industrial ML forecasting system built for Commercial & Industrial (C&I) facilities (specifically high-load steel manufacturing). It is designed to interface with Battery Energy Storage System (BESS) Energy Management Systems (EMS) to prevent utility Maximum Demand penalties.

---

Evolution Journey: MK1 to MK5

Iteration	Core Focus	Major Discovery / Bottleneck	Resolution in Next Version
MK1	Naive BESS dispatch on raw kWh	Discharged on kWh without power factor knowledge; resulted in ₹0 demand savings.	Made dispatch kVA-aware.
MK2	kVA-aware BESS Peak Shaving	Saved ₹44,068 by incorporating power factor, but used static split and idealized battery logic.	Moved to rolling ML evaluation.
MK3	Daily Walk-Forward Quantile Regressor	Blind Spots Identified: Data leakage across midnight, high false alarms (precision 14%), and the power factor conversion trap ($kVA = \text{kWh} / pf_{current}$).	Direct kVA prediction & leak-free seam.
MK4	Direct kVA + Persistence Gating	Fixed data leakage and kVA calculation. Proved that single-step quantile models suffer from high false alarm rates because Q90 predicts risk, not certainty.	Dual-channel multi-step architecture.
MK5 (Final)	Dual-Channel Multi-Step Engine	The Production Solution: 4-step forward trajectory (+15m, +30m, +45m, +60m) combined with calibrated 3-tier operational risk classification (GREEN, AMBER, RED).	Packaged into engine.py.

---

System Architecture

                       Industrial Smart Meter Telemetry
                    (kW, kVAR, Power Factor, Load Type)
                                    │
                                    ▼
                     Feature Engineering Pipeline
             (Lags: 15m..24h | Rolling Max/Mean: 1h..24h | Ramp Rates)
                                    │
                  ┌─────────────────┴─────────────────┐
                  ▼                                   ▼
        Channel 1: Trajectory Model         Channel 2: Risk Classifier
          (Multi-Output XGBRegressor)            (XGBClassifier, Calibrated)
                  │                                   │
                  ▼                                   ▼
       4-Step Demand Trajectory            1-Hour Breach Probability (%)
    [+15m, +30m, +45m, +60m] kVA            ROC-AUC: 0.933
                  └─────────────────┬─────────────────┘
                                    │
                                    ▼
                      Operational Risk Classifier
            ┌───────────────────────┼───────────────────────┐
            ▼                       ▼                       ▼
      GREEN Tier (<20%)       AMBER Tier (20-60%)     RED Tier (≥60%)
      80.7% of intervals      11.5% of intervals      7.8% of intervals
      1.2% breach risk        12.2% breach risk       51.3% PRECISION
            │                       │                       │
            ▼                       ▼                       ▼
    [SAFE_TO_CHARGE]       [HOLD_RESERVE_SOC]     [LOCK_DISCHARGE_MODE]

---

Benchmark Performance (Full-Year Backtest: Feb–Dec 2018)

1. Multi-Step Trajectory Accuracy

• +15 min Horizon: $\text{MAE} = 29.01\text{ kVA} \quad | \quad R^2 = \mathbf{0.856}$
• +30 min Horizon: $\text{MAE} = 38.86\text{ kVA} \quad | \quad R^2 = \mathbf{0.740}$
• +45 min Horizon: $\text{MAE} = 45.49\text{ kVA} \quad | \quad R^2 = \mathbf{0.672}$
• +60 min Horizon: $\text{MAE} = 53.94\text{ kVA} \quad | \quad R^2 = \mathbf{0.571}$

2. Operational Risk Tiers

• Breach Classifier ROC-AUC: $0.933$
• RED Tier (High Risk): Fires only $7.8%$ of the time with $51.3%$ Precision (eliminating operator alert fatigue).
• AMBER + RED Combined: Captures $84.5%$ of all historical demand breaches.

---

Python API Usage (engine.py)

import pandas as pd
from engine import PowerPredictorEngine

# 1. Initialize engine (loads model_trajectory.json and model_classifier.json)
engine = PowerPredictorEngine(contract_limit_kva=475.0)

# 2. Extract features from a rolling buffer of smart meter readings
recent_telemetry_df = pd.read_csv('Steel_industry_data.csv').tail(150)
features = PowerPredictorEngine.extract_features_from_buffer(recent_telemetry_df)

# 3. Generate 1-hour forecast and EMS guidance
forecast = engine.predict(features)
print(forecast)