An Adaptive Wind Power Prediction Model Based on Cross-Attention Mechanism With Incremental Alignment

Wind energy is a clean energy source that can effectively reduce pollution. Efficient and accurate wind power forecasting helps optimize the operation and scheduling of wind farms, thereby enhancing energy utilization. The intensity of wind energy is closely related to weather information, and these data are in the form of time series. Currently, most time series forecasting models use multi-head self-attention mechanisms and gated recurrent units. However, the self-attention mechanism has disorderly characteristics, which may disrupt the data sequence and lead to the loss of temporal information; meanwhile, gated recurrent units have limited effectiveness in capturing information among multiple variables, making it challenging to significantly improve prediction accuracy. To address these shortcomings, this study innovatively introduces a cross-attention mechanism into the core prediction algorithm and proposes a new forecasting module, the CA module, which can capture information among variables while maintaining the orderliness of the time series. Furthermore, to address the time lag between weather information and power data, this study designs a novel data alignment algorithm based on increments and combines it with variational mode decomposition to achieve multi-source feature alignment, further improving prediction accuracy. Additionally, a reinforcement learning algorithm is used as a hyperparameter optimization framework during the forecasting process to determine the optimal hyperparameters for different tasks, thus enhancing the model’s accuracy and practicality. The experiments are based on weather and corresponding power data from four wind power stations in eastern China during spring, including ablation, comparison, and generalization experiments. The results show that the proposed model improves wind power output forecasting accuracy by more than 15% across different time scales compared to current mainstream algorithms, with short- to medium-term prediction accuracy improvements of up to 58.1%. Note to Practitioners—The motivation of this paper is to improve the accuracy of wind power generation forecasting, thereby facilitating better regulation and dispatch of wind farms and enhancing resource utilization. Most existing forecasting models are based on methods such as Transformer and LSTM. However, the disorderly nature of these models and their limited ability to capture information among multiple variables constrain their performance, making it difficult to achieve breakthrough improvements. This paper proposes a novel forecasting module based on cross-attention. Initially, the raw data is decomposed into different frequency modes, and then the similarity between power and weather factors at different frequencies is calculated to assign different weights, enabling the model to fully capture the correlation information among multiple variables. Additionally, this paper introduces an incremental data alignment algorithm, which combines the multi-step average relative growth rate with an exponential weighted average to construct a new curve, which is used for dynamic time warping. This significantly improves alignment and eliminates the time lag effect of weather factors on power. The experimental results show that the proposed solution significantly improves prediction accuracy compared to current mainstream models. However, due to the potential influence and similarity between different variables and frequency bands, future research will explore using cross-band forecasting to address this issue. Moreover, this model can also be applied to other multi-dimensional time series forecasting tasks, such as weather forecasting, traffic congestion, and generator oil temperature prediction, demonstrating high practical value.