计及需求响应和深度结构多任务学习的电力系统短期负荷预测

马天男; 王超; 彭丽霖; 郭小帆; 付明

马天男,王超,彭丽霖,郭小帆,付明.计及需求响应和深度结构多任务学习的电力系统短期负荷预测[J].电测与仪表,2019,56(16):50-60.
matiannan,wangchao,penglilin,guoxiaofan,fuming.Power System Short-term Load Forecasting Considering Demand Response and Multi-Task Learning based on Deep Structure[J].Electrical Measurement & Instrumentation,2019,56(16):50-60.

计及需求响应和深度结构多任务学习的电力系统短期负荷预测

Power System Short-term Load Forecasting Considering Demand Response and Multi-Task Learning based on Deep Structure

DOI：10.19753/j.issn1001-1390.2019.016.009

中文关键词: 需求响应电力负荷预测深度学习 Spark平台

英文关键词:Demand Response Power load forecasting Deep learning Spark platform

基金项目:国家自然科学基金资助项目(71501071)

作者中文名	作者英文名	单位
马天男	matiannan	国网四川省电力公司经济技术研究院
王超	wangchao	国网四川省电力公司经济技术研究院
彭丽霖	penglilin	四川电力交易中心有限公司
郭小帆	guoxiaofan	国网四川省电力公司经济技术研究院
付明	fuming	北京华联电力工程监理有限公司

摘要点击次数: 2014

中文摘要:

随着需求响应技术的快速发展,使得电力系统负荷数据呈现出规模庞大、结构复杂的非线性特征,基于深度机器学习和高效数据处理平台的负荷预测方法是当前的研究重点。为实现计及需求响应的电力系统短期负荷预测,建立了基于Spark平台和时钟频率驱动循环神经网络(CW-RNNs)的短期负荷预测方法。首先,在Spark平台上设置不同工作组将全部数据分割为多个子数据模块,通过并行化计算提高数据处理效率,进而基于需求响应技术对负荷曲线做出调整,计算得到用户预期收益和用户舒适度影响指标值；其次,采用离散小波变换将调整后的负荷曲线分解,得到一组高、低频信号；并采用偏最小二乘回归模型和CW-RNNs回归模型分别对低、高频信号进行训练学习；最后,将训练好的PLS模型和CW-RNNs模型通过加权平均得到最终组合预测模型(Spark-CW-RNNs)。通过实例计算验证算法的准确性和有效性,结果表明：Spark-CW-RNNs模型比其他单一模型的预测误差更小、预测精度更高,模型具有有效性和可行性。

英文摘要:

With the rapid development of demand response technology, the load data of power system has been presented with large scale and complex nonlinear characteristics. And the load forecasting method based on deep learning and efficient data processing platform becomes the current research focus. Based on the Spark processing platform and the clock recurrent neural network(CW-RNNs), the short-term load forecasting method is established to realize the prediction of the ultra-short term load of the integrated energy systems with the demand response. Firstly, different working groups are set up on the Spark platform to divide all the data into multiple sub-data modules. The data processing efficiency is improved by the parallelization calculation, and the load curve is adjusted based on the dmand response to calculate the expected benefits and user comfort index value. Secondly, the discrete wavelet transform in used to decompose the adjusted load curve to obtain a set of high and low frequency signals. The low and high frequency signals are trained by the partial least squares regression model and CW-RNNs regression model respectively. Finally, the well-trained PLS model and the CW-RNNs model are combined to obtain the final combination forecasting model(Spark-CW-RNNs). The results show that the forecasting errors of Spark-CW-RNNs model is smaller than that of other single models, and the prediction accuracy is higher. The proposed ultra-short term load forecasting model is effective and feasible.

查看全文查看/发表评论下载PDF阅读器

关闭