Delta-Based Method Improves Electricity Load Forecasting Accuracy

This research explores a novel approach to short-term electricity load forecasting by reformulating the prediction target. Instead of directly forecasting absolute load values, the proposed method, inspired by classical time-series differencing, trains deep learning models to predict the *change* in load between consecutive time steps. The final load forecast is then reconstructed using the last observed load. This delta-based reformulation aims to stabilize the learning target and simplify the forecasting task, especially given the non-stationary nature of electricity demand influenced by weather, calendar effects, and consumption patterns. The study evaluated LSTM and Transformer models, alongside LightGBM, using real-world hourly electricity data from India. Results indicate that the delta-based approach consistently enhances hour-ahead forecasting accuracy across all models, achieving over 50% reduction in Mean Absolute Percentage Error (MAPE). For day-ahead forecasting, this method particularly benefits deep sequence models, demonstrating its power as an inductive bias for neural networks, though its effectiveness can depend on the specific model and prediction horizon.