张佐忠, 高徳民, 王浩宇, 等. 地表细小死可燃物小时步长的含水率预测模型[J]. 西南林业大学学报(自然科学), 2024, 44(5): 147–156. DOI: 10.11929/j.swfu.202307023
引用本文: 张佐忠, 高徳民, 王浩宇, 等. 地表细小死可燃物小时步长的含水率预测模型[J]. 西南林业大学学报(自然科学), 2024, 44(5): 147–156. DOI: 10.11929/j.swfu.202307023
Zhang Zuozhong, Gao Demin, Wang Haoyu, Niu Haifeng, Guo Zaijun. Moisture Content Prediction Model for Hourly Steps of Small Dead Combustibles on the Surface[J]. Journal of Southwest Forestry University, 2024, 44(5): 147-156. DOI: 10.11929/j.swfu.202307023
Citation: Zhang Zuozhong, Gao Demin, Wang Haoyu, Niu Haifeng, Guo Zaijun. Moisture Content Prediction Model for Hourly Steps of Small Dead Combustibles on the Surface[J]. Journal of Southwest Forestry University, 2024, 44(5): 147-156. DOI: 10.11929/j.swfu.202307023

地表细小死可燃物小时步长的含水率预测模型

Moisture Content Prediction Model for Hourly Steps of Small Dead Combustibles on the Surface

  • 摘要: 以张家口市崇礼区的白桦和落叶松林地为实验样地,使用传统直接估计法和长短期记忆神经网络模型(LSTM)对含水率进行单步预测,结合重新构建的直接估计法、informer、LSTM完成对不同步长含水率序列的预测,分析了informer在不依赖气象要素的条件下,对2种可燃物含水率序列的预测精度。结果表明:3种含水率序列预测模型对不同步长的含水率序列预测表现有很大差异,短时间步长上直接估计法预测精度最高,长时间步长上informer模型最好,LSTM模型次之。使用informer不仅解决了LSTM模型时间复杂度和模型的内存消耗高的问题,而且提高了对长时间步长含水率序列的预测精度。传统的气象要素回归以及直接估计方法对含水率的预测,必须依赖当前和历史时刻气象要素的实测值;使用深度学习方法解决多变量多步长的时间序列预测问题,实现对未来30 h含水率序列的预测,白桦林含水率预测MAE为0.2943,落叶松含水率预测MAE为0.1791,可以为林火预报提供参考依据。

     

    Abstract: Experimenting in Betula platyphylla and Larix gmelinii forests of Chongli District, Zhangjiakou City, traditional direct estimation methods and long short-term memory neural network models(LSTM) were used for single-step moisture content prediction. Combining restructured direct estimation methods, informer, and LSTM enabled predictions of moisture content sequences at different intervals. An analysis was conducted on the informer's accuracy in predicting moisture content sequences for 2 combustibles without relying on meteorological elements. Results revealed significant differences in the performance of 3 moisture content sequence prediction models for varying intervals. Direct estimation methods exhibited the highest prediction accuracy at shorter time intervals, while the informer model excelled at longer intervals, followed by the LSTM model. Utilizing the informer not only resolved high time complexity and memory consumption issues of the LSTM model but also enhanced the prediction accuracy of moisture content sequences over longer intervals. Traditional meteorological factor regressions and direct estimation methods for moisture content prediction rely on real-time values of current and historical meteorological elements. Using deep learning methods to address multi-variable and multi-step time series prediction achieved a 30-hour forecast of moisture content sequences. The B. platyphylla forest moisture content was predicted with an MAE of 0.2943, while L. gmelinii forest moisture content had an MAE of 0.1791, providing a theoretical basis for forest fire prediction.

     

/

返回文章
返回