Physics-Informed Dual-Attention BiGRU for Hourly Wind Speed Prediction

Abstract

The proper estimation of the near-term wind speed is one of the primary requirements to convert the power produced by wind into the electricity networks reliably. The current hybrid architectures which split the wind signal into sub-components before prediction, in spite of their competitive error measures, are costly to process and run the risk of contamination of the temporal information at the decomposition frontiers. The proposed paper suggests Physics-Informed Dual-Attention Bidirectional Gated Recurrent Unit (DA-BiGRU), which is a single-stage model designed to predict the one-hour-ahead wind speed, without explicit signal decomposition at all. Physical knowledge is added using three analytically calculated variables of the atmosphere the vertical wind shear exponent, near-surface air density, and hub-height turbulence intensity that enhance the unstructured sensor channels as structured domain-sensitive inputs. The cascaded dual-attention design is selective in the information it weights and therefore the weighted information is determined by two gates; a feature level gate that increases or decreases the relative importance of each given input variable and a time level gate which emphasizes the most predictive relevant historical instances of a 24 step lookback window. Operational record experiments of a wind farm give Mean Absolute Percentage Error (MAPE) of 5.50% and a coefficient of determination $\left(R^{2}\right)$ of 0.9873, both performing better than a naive persistence model and a rolling-decomposition LSTM benchmark. These results affirm the fact that implementing atmospheric physics into the input layer provides a computationally manageable and precise forecasting resolution that is appropriate in real-time grid application.