Short-term industrial load forecasting based on Bi-LSTM optimized by SSA and Dropout

Accurate prediction of industrial load is a crucial step in the development of smart grids. Industrial load forecasting is fundamentally a time series forecasting problem. Traditional time series forecasting models struggle to accurately predict complex industrial load changes due to their nonlinearity, time series nature, and unstationarity. Neural network models, with their robust self-learning capabilities, can effectively process industrial load data. However, these models are prone to overfitting and uncertainty issues arising from manual parameter adjustments based on experience during training. Moreover, the inability to handle bidirectional data propagation causes conventional neural network models to lose essential load data characteristics and interrelated data information. In this paper, we propose an SSA-Dropout-Bi-LSTM prediction model based on a sliding window. First, the Bi-directional Long Short-Term Memory (Bi-LSTM) neural network theory is employed to facilitate bidirectional information transfer. Dropout technique is utilized to decrease the model's overfitting degree. The sparrow search algorithm (SSA) is further used to search for the optimal hyperparameters of the Dropout-Bi-LSTM model. The model's parameter search uncertainty is reduced by dynamically adjusting parameters through the machine learning algorithm, thereby enhancing the neural network model's generalization ability. We conducted load forecasting for six industrial users in Zhejiang province. The proposed model's Mean Absolute Percentage Error (MAPE) on the test set averages at 3.75%, an improvement compared to other combinations of LSTM models (4.17%-5.37%), and a significant enhancement compared to RNN (7.22%) and GRU (5.94%). The mean coefficient of determination (R2) of the proposed model is 94.34%, which is considerably higher than RNN (87.58%) and GRU (90.28%). In comparison, the proposed model demonstrates higher prediction accuracy and better model fitting effects.