基于XGBoost特征提取的热电联产发电功率预测

Power Forecast of Cogeneration Based on Feature Extraction Using XGBoost

摘要: 精准的短期电力负载预测对电力系统的调度与控制、安全与防御等方面具有重要意义。随着电厂设备的日益复杂化，依靠专家经验或人工提取特征也变得愈发困难。结合极限梯度提升(Extreme Gradient Boosting, XGBoost)在提取特征和处理高维、非线性数据等方面的优势，提出一种基于XGBoost特征提取的融入注意力机制的深度卷积长短期记忆网络(CNN-LSTM-A)模型预测方法。首先，针对短期负荷预测，其电力负荷以及影响因素历史数据的随机性、波动性和不确定性，使用XGBoost对历史序列数据进行特征选择。将处理后的数据通过卷积层进行更深入的特征提取。然后，通过长短期记忆层进行拟合，并采取注意力机制对权值进行优化，减少历史信息的丢失并加强重要信息的影响，最终完成对功率的精准预测。利用浙江某电厂的实测数据进行实验，并与传统常用的经典模型进行对比，实验结果表明，所提方法预测精度更高，验证了该模型的精准性与可行性。

Abstract: Accurate short-term power load prediction is of great significance to power system scheduling and control, security and defense. With the increasing complexity of power plant equipment, it is increasingly difficult to rely on expert experience or manual extraction of features. Combining with Extreme Gradient Boosting (XGBoost) advantages in feature extraction and processing of high-dimensional and nonlinear data, a deep convolutional long and short-term memory network with attention mechanism (CNN-LSTM-AM) model prediction method based on XGBoost feature extraction was proposed. First of all, for short-term load forecasting, the randomness, volatility and uncertainty of historical data of power load and influencing factors, XGBoost is used to select the characteristics of historical sequence data. Further feature extraction is carried out on the processed data through the convolutional layer. Then, the long and short-term memory layer is fitted, and the attention mechanism is adopted to optimize the weights, reduce the loss of historical information and strengthen the influence of important information, and finally complete the accurate prediction of power. The measured data from a power plant in Zhejiang province are used for experiments and compared with the traditional classical model. Experimental results show that the proposed method has higher prediction accuracy, which verifies the accuracy and feasibility of the model.