采用长短期记忆深度学习模型的工业负荷短期预测方法

工业负荷不同于其他电力负荷, 受气温、时间、人口等外部因素的影响较小, 其功率需求主要由相关企业的生产计划来决定。在电力市场环境下, 准确的负荷预测有助于工业用户更好地制定电力交易策略, 从而增加收益。在此背景下, 基于改进的长短期记忆（long short term memory, LSTM）深度学习网络模型, 提出了一种工业负荷短期预测算法。首先,在网络层次上构建层数更多即网络层次更深的LSTM深度学习负荷预测模型。接着, 在每个LSTM单元构成的隐含层中, 采用Dropout技术对神经元进行随机概率失活, 并通过正则化有效避免深度学习过拟合问题并改善了模型性能。然后, 采用真实的工业用户历史负荷数据对所提算法进行测试, 并与已有的短期负荷预测算法进行对比, 包括自回归滑动平均模型 （auto-regressive and moving average model, ARMA）, 最邻近回归算法 （K nearest neighbor regression, KNN） 以及支持向量回归算法 （support vector regression, SVR）。仿真结果表明, 所提深度学习工业负荷短期预测算法相比于一些现有方法, 其预测准确度有明显提升,预测结果的平均绝对百分误差（mean absolute percentage error, MAPE）在9%以下。

A Long Short Term Memory Based Deep Learning Method for Industrial Load Forecasting

Industrial load is mainly determined by the production schedule of a given factory, with a lower correlation to external factors such as temperature, time and demography. Under electricity market environment, accurate forecasting of industrial load is essential for industry end-users to develop profitable transaction plans. Given this background, this paper studied the short-term forecasting of industrial load and proposed a long short term memory （LSTM） network based deep learning algorithm for this purpose. Compared with some existing methods, the proposed LSTM-based deep learning algorithm is improved with respect to the following aspects. First, the proposed algorithm extended the layers of the deep learning network. Thus, both the capability of the network to extract information from historical data and the capability to forecast future load are strengthened. Secondly, the dropout technique is applied to each hidden layer composed of LSTM blocks, where a probability is assigned to a hidden unit in each layer of the network. The dropout technique can prevent the neural network from overfitting through the regularization. Consequently, the overall performance of the neural network is improved. Next, actual historical data of industrial load are used to test the proposed method. Case study results show that the proposed method can significantly improve the forecasting accuracy comparing with the auto-regressive and moving average model （ARMA）, K nearest neighbor regression （KNN） and support vector regression （SVR） methods. Besides, the forecasting error measured by the mean absolute percentage error （MAPE） is less than 9% with the proposed forecasting method.