Enhanced Short-Term Wind Power Prediction via an Innovative Multitask Learning Framework Integrated With Fusion Model-Based Data Cleansing

To enhance the training efficiency and prediction accuracy of multistep short-term wind power prediction (WPP) and to tackle the challenge that data cleaning with only a single model fails to accurately detect complex outliers in wind power dataset, this study proposes a novel short-term WPP framework that integrates multitask learning (MTL) with an advanced fusion data cleaning methodology (DCM). First, a hybrid machine learning (ML) model is designed for anomaly detection and correction. The isolation forest (iForest) algorithm identifies global outliers in wind speed data, while a boundary-aware local density outlier detection algorithm captures local anomalies. Subsequently, the density-based spatial clustering of applications with noise (DBSCAN) algorithm uncovers abnormal wind speed–power generation relationships. Upon outlier detection, a backpropagation neural network performs anomaly correction, ensuring reconstructed data quality. The cleaned dataset then trains an MTL-based short-term multistep WPP model. This framework partitions different time steps into distinct tasks and introduces Gaussian perturbations to historical power values during intertask communication in training, mitigating discrepancies between training and prediction phases. A decision tree, support vector machines (SVMs), and several shallow neural networks generate preliminary multistep predictions, which are further refined by a bidirectional long short-term memory (BiLSTM) network to produce the final forecasts. Comprehensive numerical experiments validate the effectiveness of the proposed framework, demonstrating that the proposed DCM achieves 90% anomaly detection with corrected data reconstruction quality reaching $R^{2} =0.95$ , and the MTL-based WPP method outperforms existing benchmark approaches in forecasting accuracy across multiple evaluation metrics.