A Hybrid Vae–Attention Gcn–Bilstm Framework For Spatio-Temporal Weather Forecasting

Abstract

Distributed sensor networks for accurate short-range weather forecasting are difficult because of measurement noise, missing measurements, and complex spatio-temporal dependencies. In this research, an integrated four-stage deep learning system to robustly predict weather is introduced. First, a Variational Autoencoder with Noise Reduction (VAE-NR) is used to learn a 64-dimensional latent space, which is the same for the LSI space, using a symmetric autoencoder architecture, to denoise meteorological observations and to mask-conditioned reconstruct missing observations. Secondly, by introducing node-level attention and feature-level attention, a Hybrid Attention Graph Convolutional Network (HAGCN) models the relationships between stations with a Pearson-correlation based graph and mines the informative spatial representations for each station. Spatiotemporal-feature impact (STFI) mechanism then ranks and adaptively rescales the meteorological variables following the time-averaged attention scores. Third, the bidirectional long short-term memory (BiLSTM) network, which models the temporal dependence in both forward and backward directions, is used as the two-layer network. Finally, an adaptive regression boosting strategy is used to train ten learners of BiLSTM model on reweighted residual error in order to minimize systematic prediction error. The Kaggle Historical Hourly Weather data set comprises 96,453 hourly records of weather observations on eight meteorological parameters, and the experiments on this data set result in an MAE of 1.20 °C, RMSE of 1.65 °C and MAPE of 3.0%. The robustness of the proposed framework is demonstrated against the XGBoost, random forest, CNN and GCN baselines and the ablation results demonstrate that the proposed framework components complement each other.