Wind power forecasting: A hybrid multi-layer perceptron framework with adaptive noise reduction and error correction

Abstract

The increasing penetration of renewables introduces unprecedented volatility into modern power systems. Conventional forecasting frameworks often treat residual variations as unstructured noise, discarding them after correction. These approaches neglect the physical reality that residuals capture short-term disturbances, intermittency effects, and hidden fluctuations that directly affect grid stability and reliability. In this work, we propose a high-order Kalman filtering framework in which residuals are explicitly modeled as dynamic states with their own stochastic evolution. Rather than being treated as disposable errors, residuals are elevated to predictive components, enabling a simultaneous decomposition of system behavior into long-term operational trends and fast-changing renewable-driven fluctuations. The framework integrates innovation-driven covariance adaptation, allowing the filter to continuously recalibrate its process and measurement uncertainties under nonstationary grid conditions (e.g. fluctuating wind power, sudden load changes). In addition, a dual-stage neural network architecture is introduced to capture the smooth trajectory of the system state, and model high-frequency corrections. A real-time adaptive weighting strategy balances their influence, ensuring robustness both in stable operation and during disturbances triggered by renewable variability. Extensive simulations on wind power and speed datasets validate the effectiveness of the proposed method. The framework reduced mean absolute error (MAE) from 0.82 (trend based-multilayer perceptron, TMLP) and 0.88 (residual-based multilayer perceptron, RMLP) to 0.48 on the test data, representing over 40 % improvement. On the test wind speed dataset, MAE was reduced from 0.92 (TMLP) and 0.98 (RMLP) to 0.81, corresponding to gains of 14.68 % improvement.

Topics & Concepts