基于贝叶斯优化注意力机制LSTNet模型的短期电力负荷预测

Short-term Load Forecasting Based on Bayesian Optimized LSTNet with Attention Mechanism Module

为更充分挖掘多元负荷序列间的有效信息,从而提高预测精度,提出了一种集成贝叶斯超参数优化算法、注意力机制的长期和短期时间序列网络(long and short-term time-series network with attention, LSTNet-attention)以及误差修正的短期负荷预测模型。首先,构建基于贝叶斯优化的LSTNet-attention模型进行初步预测,利用贝叶斯算法优化模型多个结构参数,降低人工设置参数的随机性,并通过注意力机制合理分配特征权重;然后,通过基于贝叶斯参数优化的极端梯度提升算法(extreme gradient boosting, XGBoost)误差修正模型来挖掘初步预测误差序列中潜在、未被利用的有效信息,进行误差预测和修正,进而得到最终的预测结果。通过使用澳大利亚某地真实负荷数据进行实证分析,实验结果表明,所提预测模型相较于其他模型具有更好的预测效果,可为负荷预测等工作提供一定参考。

In order to more fully exploit the effective information among multivariate load series and thus improve the prediction accuracy,a short-term load prediction model combining Bayesian optimization algorithm,long and short-term time series network with attention mechanism and error correction was proposed.First,the LSTNet-attention model based on Bayesian optimization was constructed for preliminary prediction,and the Bayesian algorithm was used to optimize multiple structural parameters of the model,reduce the randomness of manually set parameters,and reasonably assign feature weights through the attention mechanism.Then,an XGBoost error correction model based on Bayesian hyperparameter optimization was established to mine the potential,unused and effective information in the initial prediction error sequence,and make error prediction and correction,and then obtain the final prediction results.Through the empirical analysis using the real load data of a place in Australia,the experimental results show that the proposed prediction model has better prediction effect compared with other models,which provides a relevant reference for research work such as load prediction.