基于一种新型的EMD-LSTM-ARIMA算法的超短期电力负荷预测

Ultra-Short-Term Load Forecasting Based on a New EMD--LSTM-ARIMA Algorithm

由于超短期电力负荷序列非线性和高波动性的特点,使得对其进行精确预测变得困难。因此,从滑动窗口方法的角度出发,提出了一种将EMD分解算法、LSTM神经网络以及ARIMA模型相结合的混合预测算法。所提算法首先使用EMD算法将原始电力序列分解成多个分量序列。其次,采用滑动窗口方法对分量序列进行处理,并设定窗口大小将其划分成多个窗口数据。对每个窗口数据,利用线性回归计算其决定系数,并取其决定系数平均值作为衡量该分量序列的线性趋势系数。然后,设置阈值,将线性系数与阈值进行比较,当线性系数大于阈值时,使用AR IMA模型进行预测;否则,采用LSTM神经网络进行预测。最后,将它们的预测结果相加得到所提算法的最终预测结果。为了验证所提算法的可行性与有效性,选取了第九届“中国电机工程学会杯”所提供的电力负荷数据集进行实证分析。实验结果表明了提出的算法是有效的。

Due to the nonlinearity and high volatility characteristics of the short-term electric load sequence,accurate prediction becomes challenging.Therefore,this study proposes a hybrid prediction algorithm that combines the Empirical Mode Decomposition(EMD)algorithm,Long Short-Term Memory(LSTM)neural network,and Autoregressive Integrated Moving Average(ARIMA)model,taking a perspective from the sliding window method.The proposed algorithm starts by employing the EMD algorithm to decompose the original electric load sequence into multiple component sequences.Subsequently,the sliding window method is applied to process these component sequences,dividing them into multiple window data with specified window sizes.For each window data,linear regression is used to calculate its coefficient of determination(R-squared),and the average R-squared is taken as a measure of the linear trend for the corresponding component sequence.A threshold is set to compare the linear coefficient with the threshold value.If the linear coefficient exceeds the threshold,the ARIMA model is utilized for prediction;otherwise,the LSTMneural network is employed.Finally,the prediction results from both models are aggregated to obtain the final prediction results of the proposed algorithm.To validate the feasibility and effectiveness of the proposed algorithm,this study conducted the empirical analysis using the electric load dataset provided in the 9th"China Electrical Engineering Society Cup"competition.The experimental results demonstrate the effectiveness of the proposed algorithm.