Deep Learning Model for Short-Term Island Temperature Prediction Based on Wavelet Denoising and Multi-Scale Attention Mechanism

This paper proposes an advanced deep learning model integrating Long Short-Term Memory (LSTM), wavelet denoising, and a multi-scale attention mechanism to enhance short-term temperature prediction accuracy on islands. Initially, wavelet denoising is applied to enhance data quality by eliminating noise. Subsequently, integrating a multi-scale attention mechanism into LSTM enhances the model's ability to capture temporal dependencies within sequences, thereby improving predictive performance. The model is rigorously trained and validated, and benchmarked against Nested LSTM, Attentional LSTM (ATT-LSTM), Deep GRU, CNN-LSTM, and RNN-LSTM models for 12-hour temperature forecasting. Experimental results demonstrate superior performance compared to benchmarks, with significantly lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) values, highlighting its exceptional predictive accuracy. This research provides novel insights and substantial technical advancements for enhancing the precision of island temperature predictions, contributing to the field of environmental forecasting.