Missing Wind Speed Data Prediction using Artificial Neural Network

Abstract

In southern Thailand, where monsoon-driven storms disrupt meteorological data collection, missing wind speed records hinder accurate weather forecasting and renewable energy planning. This study leverages an Artificial Neural Network (ANN) to impute missing wind speed data from the ERA5 reanalysis dataset (2012–2024) at latitude 6.0, longitude 102.0, addressing gaps that reduce forecasting precision by up to 12%. Using 114,792 hourly observations of wind speed, temperature, pressure, and precipitation, the ANN model, with 2–3 hidden layers and ReLU activation, was trained on 70% of the data, validated on 15%, and tested on 15%, incorporating 5% simulated missingness. The model achieved a Mean Absolute Error (MAE) of 0.42 m/s, Root Mean Squared Error (RMSE) of 0.58 m/s, and R² of 0.87, outperforming linear interpolation (MAE: 0.78 m/s). Feature importance analysis revealed precipitation as the dominant factor (40%), followed by temperature (25%) and pressure (20%). These results enhance storm forecasting for 2 million residents in Yala, Narathiwat, and Pattani and support Thailand’s 2,000 MW wind energy potential. Despite reduced accuracy during extreme storms (RMSE: 0.65 m/s), the ANN offers a robust solution for tropical data imputation, advancing meteorological research and sustainable energy applications. Future work should integrate real-time storm data and hybrid ANN-LSTM models to further improve predictions.