A hybrid CNN-LSTM model for accurate day-ahead wind power output forecasting with outliers detection using FCM-mahalanobis distance-ANN

Optimizing day-ahead wind power forecasts is critical for reliable grid integration of renewable energy. This study presents a hybrid model, CNN–LSTM, for high-accuracy day-ahead wind power forecasting and compares its performance against three cutting-edge models: extreme learning machine–support vector regression (ELM–SVR), support vector regression (SVR), and a standard artificial neural network (ANN). All models were trained on real-world SCADA measurements—comprising wind speed, wind direction, temperature, and power output—using a consistent seven-day sliding window. Prior to training, data underwent a uniform three-stage preprocessing pipeline: (1) fuzzy C-means clustering is employed to divide the data into different clusters, (2) Mahalanobis distance filtering to remove gross outliers, and (3) an ANN-based refinement stage for residual anomaly detection. The CNN–LSTM framework that was proposed used convolutional layers to extract spatial features and LSTM units to model temporal dependencies and achieved a mean absolute percentage error (MAPE) of 7.5%, outperforming ELM–SVR (8.4%), SVR (11.2%), and ANN (12.0%). It also had the lowest mean absolute error (MAE) of 80 kW, the lowest root mean squared error (RMSE) of 105 kW, and the highest coefficient of determination (R2) of 0.95. Furthermore, on days with extremely high or low power output, the CNN–LSTM model reached a combined accuracy of 90.1%, surpassing ELM–SVR (88.7%), SVR (80.5%), and ANN (78.9%). These results show that the hybrid model CNN–LSTM not only captures complex spatial–temporal patterns in wind datasets, but it also makes forecasts much more reliable and accurate. This makes it a great way to manage renewable energy systems. The results also demonstrate that the three-phase FCM + Mahalanobis distance + ANN preprocessing applied to a CNN–LSTM for day-ahead wind power output forecasting lowered the MAPE by ≈ 26 %, improved key error metrics by 10–12 %, and raised combined accuracy by more than 3 percentage points, highlighting its important role in making accurate predictions.