A Hybrid PSO‐CEEMDAN‐GRU Framework for Short‐Term Load Forecasting in Standalone Solar Microgrids: A Case Study in Nigeria

ABSTRACT

Accurate short‐term load forecasting is critical for efficient operation of standalone solar microgrids, especially dominated with high variability. This paper proposes a hybrid deep learning framework integrating complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), particle swarm optimisation (PSO) and gated recurrent units (GRU) for one‐step‐ahead hourly load forecasting. CEEMDAN decomposition is applied to the training set, yielding 11 intrinsic mode functions (IMFs) that isolate distinct frequency components of the load signal. An independent GRU network is trained per IMF, with PSO tuning three hyperparameters, namely, unit count, dropout rate and learning rate, independently for each IMF, producing genuinely differentiated configurations. The framework is validated on real‐world data from the Kalong solar microgrid, Niger State, Nigeria and benchmarked against LSTM, BiLSTM, GRU, BiGRU and BiLSTM‐RNN baselines. The proposed PSO‐CEEMDAN‐GRU model achieves an RMSE of 0.1415 kW, MAE of 0.1091 kW, MAPE of 2.64% and R ² of 99.05%. Ablation analysis confirms that CEEMDAN decomposition is the dominant driver with 23.0% RMSE reduction, while PSO tuning contributes an incremental gain of 12.0% when applied alone. These results demonstrate that adaptive signal decomposition with per‐component optimised recurrent modelling substantially improves forecasting accuracy for complex, non‐stationary microgrid load profiles.