Temperature Prediction for Technical Water Supply Systems in Pumped Storage Power Plants Based on Random Forest Downscaling and Improved LSTM Methods

Given the limitations of traditional pumped storage power station technical water supply systems, particularly their low level of intelligence and the challenge of predicting the coupling of numerous parameter variables, a temperature prediction model based on random forest feature dimensionality reduction and an improved LSTM neural network is proposed. The model utilizes random forest dimensionality reduction on cleaned operational data to reduce the dimensionality of high-dimensional measurement variables involved in the target prediction. This method identifies and selects the most important variables as the main features, which are then input into the LSTM neural network for prediction. The network structure is optimized using the particle swarm optimization (PSO) algorithm to establish the final temperature prediction model. The proposed random forest feature dimensionality reduction method is compared to the principal component analysis (PCA) method. The results indicate that random forest dimensionality reduction effectively removes noise and other interferences unrelated to the prediction target, while better retaining key component information. Additionally, the particle swarm optimization algorithm improves prediction accuracy to a certain degree. Overall, the constructed model accurately predicts the temperature of the technical water supply system, outperforming the traditional LSTM network in terms of prediction accuracy, and offers a valuable reference for predicting other parameters in technical water supply systems.