48-hour demand forecasting
Recursive half-hourly inference with updated lag and rolling-demand features.
Machine learning for short-term electricity demand forecasting.
An end-to-end engineering project combining NESO historical data, trained forecasting models, SHAP explainability, interactive analytics, a FastAPI backend, and a Next.js frontend.
Capabilities
Platform capabilities
The complete forecasting workflow is exposed through focused analytical tools rather than disconnected model experiments.
Forecast confidence intervals
Point-level uncertainty bands derived from saved validation error.
SHAP explainability
Local prediction contributions and global feature importance from the active model.
Interactive scenario simulation
Model-backed what-if analysis for demand, generation, weather, and calendar assumptions.
Regional demand analytics
Great Britain regional layers for demand, stress, renewables, and carbon intensity.
REST API backend
Typed FastAPI endpoints for forecasts, model metadata, simulation, and explainability.
Product
One system, organised by engineering capability.
Forecast production, operational interpretation, and model explanation remain connected to the same saved artifacts and API contracts.
Forecasting
From historical demand to an operational forecast.
A continuous 48-hour demand horizon built from half-hourly NESO observations and recursively updated model features.
48-hour prediction
Historical comparison
Peak demand detection
Confidence intervals
Operations
Demand signals organised for grid-level decisions.
Operational views connect national forecasts, regional demand allocation, scenario assumptions, and model confidence.
Grid signals
Regional map
Operational monitoring
Scenario simulation
Explainability
Model behaviour made inspectable.
Saved validation metrics and SHAP contributions show which engineered features influence both the model globally and an individual forecast.
SHAP values
Feature importance
Forecast drivers
Model insights
Architecture
A traceable path from raw data to an interactive forecast.
Each stage has a clear responsibility, typed boundary, and output that can be inspected independently.
NESO Dataset
Half-hourly demand
Feature Engineering
Lags, rolling, time
Model Training
Baseline estimators
Model Evaluation
MAE, RMSE, MAPE, R²
FastAPI API
Typed inference services
Next.js Dashboard
Interactive analytics
Technology stack
Python
FastAPI
scikit-learn
SHAP
Pandas
NumPy
Next.js
TypeScript
Tailwind
Recharts
Machine Learning
Measured as an engineering system.
The platform reports dataset scale, model performance, forecast cadence, and explainability as first-class operational facts.
17,232
Historical observations
30 min
Sampling interval
48 hrs
Forecast horizon
1.36%
Validation MAPE
11
Engineered features
Real
SHAP explainability
Live
REST API
Interactive
Scenario engine
Documentation
Follow the implementation, not just the interface.
The repository documents the architecture, machine-learning workflow, API services, and known model limitations.