Abstract
<jats:p>Accurate power load forecasting and dynamic adjustment are critical for improving energy efficiency and reducing electricity costs in energy-intensive enterprises under integrated energy system environments. To address the problems of complex computation, parameter randomness, and limited prediction accuracy in traditional load forecasting methods, this study proposes a reinforcement learning-based power load forecasting and cost-aware optimization approach. An improved particle swarm optimization (IPSO) algorithm is introduced to adaptively optimize key hyperparameters of a Long Short-Term Memory (LSTM) network, while an error-following randomized forgetting gate mechanism is designed to enhance the LSTM’s ability to track historical prediction deviations. In addition, the ReLU activation function is adopted to mitigate gradient vanishing and improve learning stability. Simulation experiments are conducted using 2,080 samples constructed from enterprise historical load data combined with seasonal and meteorological factors. Comparative results demonstrate that the proposed IPSO-EFFG-LSTM model significantly outperforms conventional LSTM and EFFG-LSTM models, achieving reductions in MSE from 685.3 to 107.28 and lowering the average prediction error to 5.2%. Furthermore, when applied to enterprise production scheduling, the optimized load forecasts contribute to a reduction in electricity costs from 712.08 million to 663.52 million, corresponding to a cost-saving rate of 6.8%. These results verify the effectiveness of the proposed method in enhancing load prediction accuracy and supporting energy-efficient decision-making.</jats:p>