Day-Ahead Forecasting using LSTM with rolling windows

The project involves day-ahead forecasting which takes a model fit into historical data and predicts future demand of electricity. Time series forecasting methods are used as the dataset contains an explicit order dependence between observations, i.e. a time dimension which makes the problem difficult to handle but makes the prediction more accurate. Various machine learning and deep learning algorithms were used to accomplish this work. Results obtained by autoregressive integrated moving average(ARIMA), autoregressive moving average with explanatory variable(ARIMAX), seasonal autoregressive moving average(SARIMA), seasonal autoregressive moving average(SARIMAX) and long short term memory(LSTM) recurrent neural networks with rolling windows were compared and the one with least error, i.e. LSTM with rolling windows was used for time series prediction.本项目旨在开展日前电力负荷预测，即通过拟合历史数据构建模型以预测未来电力需求。鉴于该数据集的观测值间存在明确的序列依赖关系（即带有时间维度），此类依赖虽提升了问题处理的复杂度，却可有效优化预测精度，因此本研究选用时间序列预测方法开展任务。本研究采用多种机器学习与深度学习算法完成该工作：对自回归积分滑动平均（autoregressive integrated moving average, ARIMA）、带解释变量的自回归滑动平均（autoregressive moving average with explanatory variable, ARIMAX）、季节性自回归滑动平均（seasonal autoregressive moving average, SARIMA）、带解释变量的季节性自回归滑动平均（seasonal autoregressive moving average with explanatory variable, SARIMAX）以及带滚动窗口的长短期记忆（long short term memory, LSTM）循环神经网络的预测结果进行对比后，最终选取误差最小的带滚动窗口LSTM模型用于时序预测任务。