Electricity demand forecasting has been a real challenge for power system scheduling in different levels of energy sectors. Various computational intelligence techniques and methodologies have been employed in the electricity market for short-term load forecasting, although scant evidence is available about the feasibility of these methods considering the type of data and other potential factors. This work introduces several scientific, technical rationales behind short-term load forecasting methodologies based on works of previous researchers in the energy field. Fundamental benefits and drawbacks of these methods are discussed to represent the efficiency of each approach in various circumstances. Finally, a hybrid strategy is proposed.

针对输入数据特征多时负荷预测模型精度提升难的问题,文章提出一种并行多模型融合的混合神经网络超短期负荷预测方法。将卷积神经网络(convolutional neural network,CNN)与门控循环单元神经网络(gated recurrent unit neural network,GRU-NN)并行,分别提取局部特征与时序特征,将2个网络结构的输出拼接并输入深度神经网络(deep neural network,DNN),由DNN进行超短期负荷预测。最后应用负荷与温度数据进行预测实验,结果表明相比于GRU-NN网络结构、长短期记忆(long short term memory,LSTM)网络结构、串行CNN-LSTM网络结构与串行CNN-GRU网络结构,所提方法具有更好的预测性能。

For the purpose of addressing the difficulty of improving load forecasting accuracy brought by enormous input data features, a method based on hybrid neural network using parallel multi-model combination is proposed. In order to respectively extract local features and time-series features, this paper places the convolutional neural network (CNN) in parallel with the gated recurrent unit (GRU) structure, then concatenates the output of two network structures and inputs to a deep neural network, uses deep neural network to perform load forecasting. Through a prediction experiment of load and temperature data by using the proposed method, the experiment results show that, compared with GRU-NN model, long short term memory (LSTM) model, serial CNN-LSTM network model and serial CNN-GRU network model, the proposed method shows better prediction performance.

相对于系统级负荷,用户负荷具有基数小、波动性与随机性更强的特点,加大了用户负荷预测的难度。文章借助互信息与深度学习理论,提出了一种基于最大相关最小冗余(max-relevance and min-redundancy, mRMR)和长短期记忆网络(long-short term memory networks, LSTM)的用户负荷短期预测模型。首先,采用mRMR算法对特征变量进行排序并选取合适的输入变量集合,mRMR既可以保证输入变量与目标值间互信息值最大,又使得变量间冗余性最小。接着,对选取的输入变量集合建立LSTM预测模型,LSTM能较好处理和预测延迟较长的时间序列,且不会存在梯度消失和梯度爆炸现象。最后,通过算例验证了所提算法的有效性。

Compared with the system-level load, consumer load has the characteristics of small base, stronger volatility and randomness, which increases the difficulty of consumer load forecasting. With the help of mutual information and deep learning theory, this paper proposes a short-term consumer load forecasting model based on max-relevance and min-redundancy (mRMR) and long short-term memory network (LSTM). Firstly, the mRMR algorithm is used to sort the characteristic variables and select a suitable set of input variables. mRMR can not only ensure the maximum mutual information value between the input variable and the target value, but also minimize the redundancy between the variables. Secondly, the LSTM forecasting model is established for the selected set of input variables. LSTM can better process and forecast time series with long delays, and there will be no gradient disappearance and gradient explosion. Finally, an example is used to verify the effectiveness of the algorithm in this paper.

短期电力负荷预测在电网安全运行和制定合理调度计划方面发挥着重要作用。为了提高电力负荷时间序列预测的准确度,提出了一种由完整自适应噪声集成经验模态分解(complete ensemble empirical mode decomposition with adaptive noise, CEEMDAN)和基于注意力机制的长短期记忆神经网络(long short-term memory network based on attention mechanism, LSTM-Attention)相结合的短期电力负荷预测模型。完整自适应噪声集成经验模态分解有效地将负荷时间序列分解成多个层次规律平稳的本征模态分量,并通过神经网络模型预测极大值,结合镜像延拓方法抑制边界效应,提高分解精度,同时基于注意力机制的长短期记忆神经网络自适应地提取电力负荷数据输入特征并分配权重进行预测,最后各预测模态分量叠加重构后获得最终预测结果。通过不同实际电力负荷季节数据分别进行实验,并与其他电力负荷预测模型结果分析进行比较,验证了该预测方法在电力负荷预测精度方面具有更好的性能。

Short-term power load forecasting plays an important role in the safe operation of power grid and the formulation of reasonable dispatching plan. In order to improve the accuracy of power load time-series forecasting, a short-term power load forecasting model based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and short-term memory neural network based on attention mechanism (LSTM-Attention) is proposed in this paper. The complete ensemble empirical mode decomposition with adaptive noise effectively decomposes the load time series into multiple levels of regular and stable eigenmode components, and suppresses the boundary effect through the neural network model prediction maximum combined with the image continuation method to improve the decomposition accuracy. At the same time, the long short-term memory neural network based on attention mechanism adaptively extracts the input characteristics of power load data and assigns weights for prediction. Finally, the final prediction results are obtained after superposition and reconstruction of each prediction modal component. Experiments are carried out on different seasonal data of actual power load, and the results of other power load forecasting models are analyzed and compared to verify that the forecasting method has better performance in power load forecasting accuracy.

综合考虑风电功率序列及气象数据的多维特征,提出一种弹性网稀疏核主成分分析（EN-SKPCA）降维方法,对气象因素降维并表述为回归优化型问题,添加的弹性网惩罚解决了KPCA重构主成分难以解释构成的问题;提出花授粉算法（FPA）优化长短时记忆神经网络（LSTMNN）预测模型,可自动筛选出最佳超参数,降低了参数经验设置所带来的随机性。该方法解决了突变天气的影响,提高了预测精度。对2017年宁夏麻黄山第一风电场实测数据实验,证明了该方法的优越性。

Comprehensively considering the characteristics of wind power series and the multi-dimensional meteorological data, a dimensionality reduction method of elastic net improved kernel principal component analysis （EN-SKPCA） is proposed. The dimensionality of meteorological factors is reduced and expressed as a regression optimization problem. The added elastic network penalty solve the problem that the KPCA reconstruction principal component is difficult to explain. The flower pollination algorithm （FPA） is proposed to optimize the long-short-term memory neural network （LSTMNN） prediction. The model can automatically select the best super parameters and reduce the randomness caused by the empirical setting of parameters. The method solves the influence of abrupt weather and improves the prediction accuracy. The superiority of this method is proved by the experiment on the measured data of Mahuangshan No.1 wind farm in Ningxia in 2017.