Advancing long-term precipitation pattern forecasting in Atlantic Canada using successive variational mode decomposition, recursive LSTM, and graph-based feature selection

Advancing long-term precipitation pattern forecasting in Atlantic Canada using successive variational mode decomposition, recursive LSTM, and graph-based feature selection

Precipitation forecasting is crucial in Canada's Maritime provinces, given their unique geography and susceptibility to precipitation impacts. Accurate forecasts aid farmers, transportation authorities, and climate change adaptation efforts, ecosystems, and infrastructure. This study introduces a groundbreaking multi-temporal deep-learning framework for forecasting monthly precipitation in Canada's Maritime region, encompassing Charlottetown and St. John's, distinguishing it as a trailblazing innovation among cutting-edge complementary deep-learning algorithms. This pioneering research unveils, for the first time, a cutting-edge hybrid framework that synergizes successive variational mode decomposition (SVMD), recursive long short-term memory (RLSTM), graph feature selection, and Borda count-based multi-criteria decision-making (BORDA). The innovative aspect lies in the recursive architecture of RLSTM, which sets it apart from traditional SVMD-LSTM hybrids by enabling multi-horizon memory feedback loops that improve long-range temporal learning. Integrating graph-based feature selection with partial autocorrelation function (PACF) analysis enhances the extraction of the most informative SVMD components, enhancing prediction accuracy and reducing model complexity. Furthermore, the framework is distinguished by its precise and efficient performance, facilitated by intuitive hyperparameter configurations during both the decomposition and training stages. It provides a pragmatic and scalable alternative to other leading complementary deep-learning methods. To benchmark the performance of the primary model, it is compared against a convolutional neural network-long short-term memory (CNN-LSTM), random vector functional link (RVFL), and a light gradient-boosting machine (LightGBM), with a rigorous evaluation of both standalone and hybrid counterparts. To assess the accuracy of the model, a single metric comprising six statistical indices, including correlation coefficient (R), Nash–Sutcliffe efficiency (NSE), and Kling–Gupta efficiency (KGE), consolidated via BORDA, was employed to simplify the identification of superior frameworks. An accuracy assessment in Charlottetown reveals that SVMD–RLSTM, owing to optimal metrics (BORDA 0.95, R = 0.9508, and RMSE = 15.6567 mm|T + 1; BORDA = 0.7834, R = 0.9297, and RMSE = 18.7170; |T + 3, BORDA = 0.6855, R = 0.906, and RMSE = 21.3539|T + 7), outperformed SVMD–RVFL (BORDA|T + 1 = 0.927 and BORDA), SVMD-CNN-LSTM (BORDA|T + 1 = 0.8037), and SVMD-LightGBM (BORDA|T + 1 = 0.713); whereas a diagnostic assessment in St. John's station confirms the superiority of SVMD–RLSTM (BORDA = 0.9171, R = 0.9337, and RMSE = 19.0093 mm|T + 1; BORDA = 0.5951, R = 0.9251, and RMSE = 23.6887; |T + 3, BORDA = 0.6898, R = 0.9157, and RMSE = 21.3299 mm|T + 7) over the other hybrid frameworks.