基于滑动窗口的LSTM地温预测方法

LSTM ground temperature prediction method based on sliding window

为了连续预测未来多日的日平均地温,采用LSTM和滑动窗口(SW)相结合的方法,以7天为周期更新输入参数,实现连续预测。分析成都市近50年的地温数据的关系发现,日平均地温与前几日平均地温呈现较强的正相关关系,但在夏季这种关系较弱。进一步根据季节特点建立了5个预测模型,分别为:全年、春季、夏季、秋季和冬季。结果表明,冬季模型预测效果最佳,平均绝对误差约为0.2641℃;春秋两季模型预测结果的平均绝对误差接近,分别为0.3867℃和0.4064℃;夏季模型预测结果的平均绝对误差略高,约为0.7516℃;全年模型预测结果的平均绝对误差约为0.6546℃。研究发现LSTM模型对于有明显递增或递减趋势的时间序列,训练和预测效果更好。通过对比,周期性SW+LSTM方法比传统LSTM和BP-LSTM方法的预测效果更好。

In order to continuously predict the daily average ground temperature in the future days,the method of combining LSTM and SW is adopted to update the input parameters periodically in 7 days,and the effect of continuous prediction is achieved.Correlation analysis on the recent 50 years ground temperature data in Chengdu reveals that,except in summer season,there is a strong positive correlation between the daily average ground temperature and the last few days average ground temperature.According to the seasonal characteristics,five prediction models,including the whole year,spring,summer,autumn and winter,are established.The results show that the prediction effect of winter model is the best,with the average absolute error of about 0.2641℃.The mean absolute errors of the forecast results in spring and autumn are 0.3867℃and 0.4064℃,respectively.The mean absolute error of summer model prediction is slightly higher,about 0.7516℃.The average absolute error of the annual model prediction is about 0.6546℃.It is found that LSTM model has better training and prediction effect for time series with obvious increasing or decreasing trend.By comparison,the prediction effect of periodic SW+LSTM method is better than traditional LSTM and BP-LSTM method.SW+LSTM realizes the continuous prediction of the ground temperature for several days in the future with a low error,and can provide guiding suggestions for the research work of the future ground temperature prediction.