Grid Energy Intelligence
Energy demand fluctuates with weather, seasonality, and economic activity. Grid operators today rely on deterministic forecasts that provide no insight into prediction uncertainty, forcing costly reserve margins. Our multi-horizon model captures seasonal patterns across all Indian states and outputs calibrated prediction intervals, enabling operators to optimise reserve allocation with quantified confidence.
State-Level Forecast Explorer
Explore energy demand and peak load forecasts across 32 Indian states. Click a state on the map or use the dropdown to view its forecast.
Karnataka
Temporal Fusion Transformer
Our architecture uses a multi-horizon attention mechanism over calendar features and historical load data, combined with quantile output heads for probabilistic forecasting.
Multi-Source Fusion
Fuses historical load data and calendar features onto a unified temporal embedding with state-level context.
Temporal Attention
Variable selection network weights each exogenous input per timestep, learning which factors matter when.
Quantile Output Heads
Separate decoders for energy and peak demand predict 10th, 50th, and 90th percentiles with pinball loss.
Static Covariate Encoding
State-level embeddings and region identifiers allow the model to generalize across all 32 Indian states.
32 States, 2 Targets
Trained on 72 days of context to predict 24 days ahead across all 32 Indian states. The model uses a Transformer architecture with AdamW optimiser, cosine decay, and early stopping.
Dataset Composition
32 Indian states with 9,430 test samples. Two target variables: daily energy demand (MU) and peak load (MW). 72-day sliding context windows.
Training Configuration
AdamW optimiser with cosine annealing, 4-layer Transformer with 4 attention heads, 256 hidden dimensions, trained with pinball loss across 3 quantiles.
Context & Prediction Windows
72 days of historical context to forecast 24 days ahead. Quantile outputs at P10, P50, and P90 for probabilistic energy and peak demand predictions.
Performance Highlights
R² of 0.984 for both energy and peak across all states. PICP of 90.4% for energy and 89% for peak against a 90% target interval.
Benchmarking accuracy
Unified model achieves R² = 0.984 for both energy demand and peak load across 32 states with 90.4% prediction interval coverage using a single architecture that generalizes across diverse consumption patterns.
Average daily energy demand error
Variance explained for energy demand
Prediction interval coverage (energy)
Average uncertainty band width
Average peak demand error
Variance explained for peak load
Prediction interval coverage (peak)
Average uncertainty band width
Across India
Root mean squared error
Root mean squared error
Symmetric MAPE
Horizon & seasonal patterns
R² stays above 0.99 at day+1 and degrades gradually to 0.98 by day+24. Seasonal errors show minimal variation while the model handles summer peaks and monsoon troughs equally well.
Incorporate predictions
API access for pilot partners is planned as the platform matures.
curl -X POST https://api.purnabodha.ai/v1/predict/energy \
-H "Authorization: Bearer $API_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"state": "Karnataka",
"context_days": 72,
"quantiles": [0.10, 0.50, 0.90]
}'Where We're Taking This Next
Every deployed system raises new questions. These research programs explore the next generation of capabilities for grid energy intelligence.
Renewable Integration Forecasting
Predicting demand under high renewable penetration with intermittent generation patterns from solar and wind sources.
Extreme Event Forecasting
Rare but critical grid events including heatwaves, cold snaps, and unplanned outages that challenge conventional load forecasting.
Cross-Region Transfer Learning
Adapting models trained on data-rich states to data-sparse regions with different consumption patterns and infrastructure.
Real-Time Adaptive Recalibration
Continuously updating prediction intervals as new smart meter data streams arrive, maintaining calibration under changing grid conditions.
Interested in piloting this on your grid?
We're looking for state load dispatch centres and renewable energy parks interested in testing this model on real grid data.