Abstract:Gas load forecasting is affected by various complex factors such as social economy, weather factors, date types, and the combination of multiple factors, and it will inevitably lead to a large randomness and a certain degree of complexity in the trend of gas load sequence changes. In order to effectively improve the accuracy of gas load forecasting, a new integrated deep learning algorithms is proposed to predict the gas load in multiple steps. Firstly, the non-stationary nonlinear load sequence is decomposed into several steady-state and linear IMF components and residuals by the set of EEMD algorithm, which effectively avoids the modal aliasing problem caused by the traditional EMD. Then, each subsequence obtained by EEMD decomposition is composed of a training matrix different from the feature sequence extracted by AutoEncoder. After that, each subsequence obtained by EEMD decomposition is composed of a training matrix different from the feature sequence extracted by AutoEncoder. Finally, the corresponding Long Short Term Memory (LSTM) prediction model is established for the training matrices corresponding to different subsequences, and the component prediction values are reconstructed to obtain the final prediction result. In order to verify the effectiveness and prediction performance of the proposed algorithm, the Shanghai gas data was used to simulate the above model. The results show that the prediction accuracy is significantly improved compared with the comparison method.