Title:FATE: Focal-modulated Attention Encoder for Temperature Prediction

Abstract:One of the major challenges of the twenty-first century is climate change, evidenced by rising sea levels, melting glaciers, and increased storm frequency. Accurate temperature forecasting is vital for understanding and mitigating these impacts. Traditional data-driven models often use recurrent neural networks (RNNs) but face limitations in parallelization, especially with longer sequences. To address this, we introduce a novel approach based on the FocalNet Transformer architecture. Our Focal modulation Attention Encoder (FATE) framework operates in a multi-tensor format, utilizing tensorized modulation to capture spatial and temporal nuances in meteorological data. Comparative evaluations against existing transformer encoders, 3D CNNs, LSTM, and ConvLSTM models show that FATE excels at identifying complex patterns in temperature data. Additionally, we present a new labeled dataset, the Climate Change Parameter dataset (CCPD), containing 40 years of data from Jammu and Kashmir on seven climate-related parameters. Experiments with real-world temperature datasets from the USA, Canada, and Europe show accuracy improvements of 12\%, 23\%, and 28\%, respectively, over current state-of-the-art models. Our CCPD dataset also achieved a 24\% improvement in accuracy. To support reproducible research, we have released the source code and pre-trained FATE model at \href{this https URL}{this https URL}.