Learning Temporal and Spatial Correlations Jointly: A Unified Framework for Wind Speed Prediction

TLDR

Predicting wind speed by jointly exploiting temporal and spatial correlations is a challenging, underexplored problem. The study investigates multi‑site wind‑speed prediction using spatio‑temporal correlations. The authors propose PSTN, a unified CNN–LSTM architecture that extracts spatial features from wind‑speed matrices, models their temporal evolution with LSTM, and jointly learns spatial and temporal correlations end‑to‑end. Experiments for short‑term predictions on real‑world data show that PSTN outperforms prior methods.

Abstract

Leveraging both temporal and spatial correlations to predict wind speed remains one of the most challenging and less studied areas of wind speed prediction. In this paper, the problem of predicting wind speeds for multiple sites is investigated by using the spatio-temporal correlation. We proposed a deep architecture termed predictive spatio-temporal network (PSTN), which is a unified framework integrating a convolutional neural network (CNN) and a long short-term memory (LSTM). Initially, the spatial features are extracted from the spatial wind speed matrices by the CNN at the bottom of the model. Then, the LSTM captures the temporal dependencies among the spatial features extracted from contiguous time points. Finally, the predicted wind speeds are given by the last state of the top layer of the LSTM, which are generated by using the spatial features and temporal dependencies. Though composed of two kinds of architectures, PSTN is trained with one loss function in an end-to-end manner, which can learn temporal and spatial correlations jointly. Experiments for shortterm predictions are conducted on real-world data, whose results demonstrate that PSTN outperforms prior methods.