Attention Mechanism and Weighted Trend Loss for Wind Speed Correction

To address the systematic bias in simulating 10-m wind speed using MIROC6 model within CMIP6 framework, a novel approach based on deep learning is proposed. A non-stationary Informer model (Ns-Informer) is integrated with an adaptive-length attention mechanism and a multilayer perceptron (MLP) dynamic adjustment module to enhance the model's correction performance, particularly in high wind speed conditions. A new weighted trend-based mean squared error loss function is developed to optimize the correction process. This function integrates weighted error allocation with trend consistency constraints, effectively balancing the trade-off between minimizing errors at high wind speeds and preserving the temporal trends of wind speed distributions. The model utilizes a sparse attention mechanism that reduces computational complexity by concentrating on the most relevant interactions within the input data. Additionally, de-stationary factors (τ and Δ) are introduced to dynamically adjust attention weights, thereby enhancing the model's ability to capture non-stationary features, particularly high-speed anomalies. The experimental dataset includes historical simulations from 1961 to 2014, SSP1-2.6 scenario projections for years from 2015 to 2022, and CN05.1 gridded observations. Four representative stations including Beijing (moderate wind speed), Guaizihu and Mangya (high wind speed), and Ji'an (low wind speed) are selected for validation. Results indicate that Ns-Stationary Informer model significantly enhances the spatiotemporal characteristics of wind speed at both seasonal and decadal scales. The root mean square error (RMSE) of ensemble mean is reduced by 20%-50% compared to raw MIROC6 outputs. Specifically, for high wind speed stations (Guaizihu and Mangya), the RMSE is reduced by as much as 60% when wind speeds exceed 5 m·s-1. Monthly-scale wind speed trends corrected by Ns-Informer are improved in their agreement with observations. Additionally, the correction is particularly effective during summer and autumn, with the monthly-scale root mean square error being reduced by an average of over 25%. Further validation for 2015-2022 period under SSP1-2.6 scenario confirms the model's robustness, particularly in stabilizing corrections for high wind speed thresholds (greater than 5 m·s-1). This approach, by integrating non-stationary architecture optimization with dynamic loss function design, offers a new framework for correcting biases in climate models and enhances the accuracy of wind energy assessments in future climate change scenarios. Despite its advantages, this method is limited by its reliance on a single climate model (MIROC6) and a limited number of training ensemble members. Future work will expand this method to incorporate multi-model ensembles and long-term climate projections. The focus will be on further refining adaptive loss function and enhancing the model's generalization capabilities across various regions and scenarios.