Comparative Evaluation of LSTM, BiLSTM, and a Hybrid CNN-BiLSTM-Attention Model for Short-Term and Day-Ahead Electricity Load Forecasting

Abstract

Reliable electricity load forecasting is essential for grid stability, generation scheduling, and demand-supply balancing, particularly in regions such as Pakistan that face chronic supply shortfalls. Building on a base study that compared Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) models against ARIMA and SARIMA on a univariate prosumer dataset, this paper extends the comparison in three directions: it evaluates a proposed Hybrid CNN-BiLSTM-Attention architecture alongside the LSTM and Bi-LSTM baselines, it replaces the single date time feature of the base study with a fifteen-column multivariate feature set (cyclical calendar encodings, lag features at 1, 24, and 168 hours, and 24-hour rolling statistics), and it evaluates every architecture at both a 1-hour-ahead and a 24-hour-ahead forecasting horizon rather than one-step-ahead alone. All three models were trained and tested under identical conditions on the public PJM East (PJME_MW) hourly electricity load series (145,224 usable hourly records, January 2002 to August 2018) using a 168-hour lookback window, an 80/10/10 chronological train/validation/test split, the Adam optimizer, and early stopping. On the held-out test set, Bi-LSTM produced the lowest error for 1-hour-ahead forecasting (RMSE = 528.17 MW, MAE = 402.62 MW), while LSTM was marginally the most accurate model for 24-hour-ahead forecasting (RMSE = 2014.78 MW, MAE = 1475.48 MW), narrowly ahead of Bi-LSTM (RMSE = 2033.93 MW). The Hybrid CNN-BiLSTM-Attention model recorded the highest absolute error at both horizons (RMSE = 2658.79 MW at h = 1; RMSE = 2467.19 MW at h = 24), but its error changed by only -7.2% between the two horizons, compared with +103.4% for LSTM and +285.1% for Bi-LSTM, indicating markedly better cross-horizon stability. Attention-weight analysis showed the hybrid model consistently assigning the largest weight to the most recent hour and a secondary peak at the 168-hour (one-week) lag, giving a transparent, interpretable account of which historical hours’ drive each forecast. The results suggest that, under the tested configuration, richly engineered multivariate features contribute more to forecasting accuracy than additional architectural complexity, while attention-based hybrid models trade some absolute accuracy for stability and interpretability across forecast horizons.