Effective LSTMs with seasonal-trend decomposition and adaptive learning and niching-based backtracking search algorithm for time series forecasting

Long short-term memory faces challenges in information mining and parameter selection due to inherent uncertainty and randomness. In this study, we propose a novel hybrid model called 'STL-ALN_BSA-LSTM' that aims to tackle this issue. In the proposed model, data is first prepossessed using seasonal-trend decomposition to extract informative features. Next, we propose an adaptive learning and niching-based backtracking search algorithm to enhance the robustness and learning ability of the original backtracking search algorithm, which affords a balance between exploration and exploitation. Subsequently, the weights and thresholds within each LSTM are optimized by our optimization algorithm automatically, leading to improved interpretability and accuracy of the model. Finally, the proposed optimization algorithm was evaluated using five benchmark functions, and the hybrid model was compared with sixteen baselines on two real-world datasets. For rainfall forecasting, the hybrid model achieved the root mean square error of 7.3772, mean absolute error of 2.6187, and R2 of 0.8388. For power generation forecasting, our model outperformed the second-best model by 29.6%, 34.4%, and 16 ‰, respectively, as in the above metrics. These results demonstrate that the proposed hybrid model can attain higher accuracy in real-life applications than previous models, which contributes in-depth insights into parameter optimization for time series forecasting, and provides valuable suggestions in resource planning, risk reduction, and decision-making. The core code of this work can be found from https://github.com/zjuml/STL-ALNBSA-LSTM.