基于历史趋势和预测误差的时序序列预测方法研究

Research on time series forecasting method based on historical trend and forecast error

针对时间序列预测方法训练复杂、迁移预测能力差等问题,提出一种自适应预测方法。先根据预测误差和当前值确定下一时刻预测值的范围,再结合短期历史趋势确定最终预测值。得到的当前预测值代入下一轮循环中继续预测,通过不断“预测-校正-预测”循环实现对数据预测。最后利用金融、风力等时序数据,LSTM、SVM、ARIMA、MA等经典时间序列预测算法在预测精度、迁移预测能力、运算速度等方面做了对比。

Aiming at the problems of complex training and poor transfer forecasting ability of time series forecasting method,an adaptive forecasting method is proposed.The range of the next forecast value is first determined according to the forecast error and the current value,and then the final forecast value is determined by combining the short-term historical trend.The obtained forecast value is then substituted into the next cycle to continue forecasting,and the data forecasting is achieved through a continuous"forecast-correction-forecast"cycle.Using the time series data such as financial and wind power,the proposed algorithm is compared with LSTM,SVM,ARIMA,MA and other classical time series forecasting algorithms in terms of forecasting accuracy,transfer forecasting ability and computing speed.