<p>This study investigated the effectiveness of SWT data decomposition in enhancing CNN-GRU early-, slow- and late fusion models for 14-day ahead water level prediction at the Klang Gates Dam. The impact of replacing CNN in the CNN-GRU fusion models with multi-scale CNN modules, was also investigated. Multivariate time series, including daily rainfall, evaporation, monthly water demand, ONI, SOI and DMI, were used together with daily water level time series. The results demonstrated that applying SWT to decompose daily water level, rainfall and evaporation data improved the performance of all CNN-GRU fusion models. Further ablation test showed that the decomposition of water level contributed the most improvement on the models. Meanwhile, utilizing multi-scale CNN module in CNN-GRU fusion models had also enhanced the model performance. Using three multi-scale CNN modules for multivariate feature input channel in the slow and late fusion model had improved the model performance further. The improvement was attributed to the ability of the multi-scale CNN modules in providing features with various kernel sizes. The CNN-GRU slow fusion model with multi-scale CNN modules trained with SWT decomposed dataset achieved the best performance, with NRMSE, NSE and MAPE of 0.4144, 0.7919 and 0.318, respectively, for the 14-day ahead prediction.</p>