Addressing Weight Competition in Long Short-Term Memory (LSTM) Networks for Improving Temperature Prediction

This conference paper presents a study on the application of Long Short-Term Memory (LSTM) networks to predict temperature changes in the upcoming week. During our research, we encountered the issue of weight competition within the LSTM architecture, which can lead to suboptimal predictions. To overcome this challenge, three novel methods including Independent Time Series (ITS) Model, Arithmetic Segmented Time Series (ASTS) Model and Arithmetic Segmented Rolling Time Series (ASRTS) Model, respectively, were used to address weight competition and enhance the accuracy of temperature change forecasts. The effectiveness of these methods were evaluated through extensive experiments and compared against standard LSTM models. The results demonstrate the significance of mitigating weight competition to achieve more reliable and accurate temperature predictions.