基于模型无关优化策略的风电短时回归预测框架设计

目前，风电出力预测面临跨环境、跨传感器设备的多任务挑战，往往需要对不同的预测目标各自独立地展开针对性训练。鉴于此，首先提出了一种基于模型无关元学习策略 (model-agnostic meta-learning, MAML)的短期预测方法，并基于该方法能够实现对新任务样本快速适应的能力设计了新型回归训练框架。然后结合卷积神经网络–长短期记忆网络、有注意力机制的Seq2Seq、有自注意力机制的Transformer、Synthesizer等时序深度网络模型，将该框架应用于风力发电预测场景。实验结果表明相比常规的预训练–微调的深度网络训练方法，所提出的方法在GEFCom2012数据集上对各算例实现了均方根误差和均方误差等指标的提高，同时各模型在短时风电出力为案例的预测任务上的泛化性能获得了一定提升。该训练框架可便捷地将主流深度学习模型和数据集转换为适应MAML策略的匹配模式。

A Regression Framework Design for Short Term Forecasting of Wind Power Based on Model-Agnostic Meta-Learning Strategy

At present, the wind power output prediction has to face with the multi-task challenges including cross environment and cross transducer equipment, so it often needs to conduct targeted training independently for different prediction targets. For this reason, firstly, a short-term prediction method based on model-agnostic meta-learning (abbr. MAML) was proposed. Secondly, based on the ability of the proposed method, by which the new task samples could be rapidly adapted, a new regression training framework was designed. Thirdly, combining with such sequential depth network models as the convolutional neural network-long and short term memory networks (abbr. CNN-LSTM), the Seq2Seq enhanced with the attention mechanism, the Transformer and Synthesizer enhanced with self-attention mechanism, this framework was applied to the wind power forecasting scene. Experiment results show that comparing with conventional pre-training-fine-tuning deep network training method, the proposed method improves such indicators as root-mean-square error (RMSE) and mean square error (abbr. MSE) on the dataset GEFCom2012 for each computing example, meanwhile, the generalization performance of each model on the prediction task, which takes short-term wind power output as the case, obtains a certain improvement. Besides, this training framework can easily convert the mainstream deep learning regression model and its dataset to the matched pattern adapted to model-agnostic meta-learning (abbr. MAML) strategy .