How Gridalytics Works

All demand values in Gridalytics are shown in MW (Megawatts) — this is the standard unit used by the Delhi SLDC (State Load Despatch Centre) and the Indian power grid.

All three prediction types show demand in MW, but at different time scales:

All demand data comes from the Delhi State Load Despatch Centre (SLDC), operated by Delhi Transco Limited (DTL). The SLDC monitors and controls the Delhi power grid in real-time.

Delhi's power is distributed by 5 DISCOMs (distribution companies):

Gridalytics uses multiple machine learning models trained on 93 engineered features derived from demand history, weather, holidays, time patterns, and seasonal cycles.

MAPE (Mean Absolute Percentage Error) is the primary accuracy metric. A MAPE of 0.52% means the model's predictions are, on average, within 0.52% of the actual demand. For a 4,000 MW demand, that's an error of just ~21 MW.

Every day, Gridalytics automatically:

1. Predicts tomorrow's demand and records the prediction
2. The next day, fetches the actual demand from SLDC
3. Computes the error (MAPE) and logs it
4. Monitors if accuracy is degrading (drift detection)

Over the last 29 days of live tracking, the system achieved an average MAPE of 0.97% — meaning predictions were within 1% of actual demand on average.

Visit the Accuracy page to see the full predicted vs actual history, rolling MAPE trend, and drift detection status.

Delhi has some of the most extreme seasonal demand variation in the world:

The summer peak (8,500+ MW) is more than 3x the winter minimum (2,500 MW). This extreme variation is driven almost entirely by air conditioning. Delhi generates almost no power locally — all electricity is purchased from generators across Northern India.

Delhi's electricity is distributed by 5 DISCOMs (distribution companies). Gridalytics forecasts demand for each sub-region by applying historical demand ratios to the total Delhi forecast.

The ratios are computed dynamically from the last 30 days of actual SCADA data. You can see the DISCOM breakdown on the Dashboard (horizontal bar chart) and request per-region forecasts via the API.

The What-If page lets you override weather and event conditions to explore their impact on demand. This is useful for:

• Grid capacity planning — "What if we get a 48°C heatwave next week?"
• Emergency preparedness — "How much extra power do we need for Diwali?"
• Pollution impact — "If AQI hits 500 (hazardous), how does demand change?"
• Sensitivity analysis — "How much does each °C increase add to demand?"

How to interpret results: The blue line shows the baseline forecast (using actual/predicted conditions). The orange line shows your scenario. A positive percentage change means the scenario increases demand — the grid needs more power. A negative change means lower demand. Both average and peak impacts are shown separately because peak capacity is the critical safety constraint.

On the Forecast page, you can select which model to use via the Model dropdown:

• Auto (Best) — Uses the champion model for the selected resolution. Recommended for most use cases.
• LightGBM — Gradient boosting. Fastest, best for 5-minute. Optuna-tuned version available for hourly.
• XGBoost — Gradient boosting. Champion for hourly resolution.
• Ensemble — Weighted average of all loaded models. Can be more stable but not always most accurate.
• LSTM / BiLSTM — Deep learning. Captures sequence patterns but less accurate than tree models on this data.

For most users, Auto is the best choice. It automatically selects the champion model for each resolution (LightGBM for 5-min and daily, XGBoost for hourly).