Short-term natural gas load forecast algorithm based on WT+ improved SSA-LSTM model

A combined forecast model (named as the improved SSA-LSTM) that integrates the improved Sparrow Search Algorithm (SSA) with the Long Short-Term Memory (LSTM) was proposed to improve the forecast accuracy of natural gas load. In this model, high-quality initial population was produced by Sobol sequence, and then the adaptive weight and T distribution variation were introduced to increase the sparrows to jump out of local optimum and improve the global search capability. Finally, the performance of the improved SSA algorithm was verified by 12 test functions, and the combination forecast model was applied to the natural gas load forecast of a gas gate station in a city in North China. To further improve the forecast accuracy of the model, using Wavelet Transform (WT), the data on natural gas load was decomposed by five wavelets (Symlets wavelet, Coiflets wavelet, Fejer-Korovkin wavelet, Haar wavelet and Discrete-Meyer wavelet), and then put into the improved SSA-LSTM model for training and forecasting. Thus the prediction results were reconstructed and MAPE, RMSE and MAE were selected as the model evaluation indexes. The results show that the 7-level Discrete-Meyer wavelet decomposition method has the best forecast accuracy as high as 99.14%, which is 3.55% higher than the improved SSA-LSTM model and 9.31% higher than the conventional LSTM model. Therefore, this method can provide a reference for the stable supply of natural gas in the region.