Spatio-Temporal Weather Forecasting via Variational Autoencoder Denoising, Hybrid Attention Graph Convolution, and Adaptive Ensemble Bidirectional LSTM

Abstract

Accurate short-range weather forecasting over distributed sensor networks remains challenging owing to sensor noise, missing observations, and complex nonlinear spatio-temporal dependencies. This paper presents a four-stage end-to-end framework that addresses these challenges in sequence. In the first stage, a Variational Autoencoder with Noise Reduction (VAE-NR) employing a 64-dimensional latent space and a symmetrical encoder–decoder denoises raw meteorological inputs and imputes missing values by mask-conditioned reconstruction. In the second stage, a Hybrid Attention Graph Convolutional Network (HAGCN) constructs a Pearson-correlation-based inter-station graph and jointly learns node-level and feature-level attention weights to produce compact spatio-temporal representations. A Spatio-Temporal Feature Impact (STFI) algorithm then ranks and rescales input features according to time-averaged attention scores prior to sequence modeling. In the third stage, a two-layer Stacked Bidirectional Long Short-Term Memory (BiLSTM) network captures both forward and backward temporal context. In the fourth stage, regression-based Adaptive Boosting sequentially trains M = 10 BiLSTM weak learners on exponentially reweighted residuals, forming a weighted ensemble that suppresses systematic bias. Experiments on the publicly available Kaggle Historical Hourly Weather dataset—comprising 96,453 hourly records across eight meteorological variables—demonstrate that the proposed framework achieves a Mean Absolute Error of 1.20 °C, Root Mean Square Error of 1.65 °C, and Mean Absolute Percentage Error of 3.0%, surpassing XGBoost, Random Forest, CNN, and GCN baselines. Ablation analysis confirms that each component contributes independently to the final accuracy gain.