Abstract:

For dealing with the forest fire data that inhabits the properties of high dimension， variability and few useful data， a forest fire disaster area prediction method by fusing Auto-Reservoir Neural Network （ARNN） and Long Short-Term Memory （LSTM） network was proposed. Firstly， based on the characters of the data， the weights were randomly generated by using the ARNN. Then， LSTM network was introduced to further train the wights so as to reduce the stochastic influence from the random weight as far as possible. Finally， the proposed method was implemented on the fire data of Montesinho National Forest Park in Portugal to show its efficiency and effectiveness. Experimental results show that under the evaluation of Mean Absolute Percentage Error （MAPE）， the proposed framework are more stable， more effective than Back Propagation Neural Network （BP-NN）， Convolutional Neural Network （CNN）， LSTM and ARNN， and its prediction accuracy is improved by 93.93%， 47.75%， 55.27%， 9.39%， respectively. It can be seen that the proposed ARNN-LSTM fusion prediction method can efficiently and accurately predict the disaster area of forest fire.

Key words: Auto-Reservoir Neural Network (ARNN), Long Short-Term Memory (LSTM) network, fire disaster area prediction, time series prediction, reservoir computing

摘要：

针对森林火灾数据维度高、易变性、可用信息少的问题，提出一种融合自动储备池神经网络（ARNN）与长短时记忆（LSTM）网络的森林火灾成灾面积预测方法。首先，根据森林火灾数据特征，基于ARNN随机生成权重；然后运用LSTM对随机权重进行进一步训练，以减少它的随机性带来的影响；最后以葡萄牙Montesinho国家森林公园的火灾数据集为例，验证所提ARNN-LSTM融合模型预测方法的合理性和有效性。实验结果表明，在平均百分比误差（MAPE）统计指标下，与目前流行的反向传播神经网络（BP-NN）、卷积神经网络（CNN）、LSTM和ARNN预测方法相比，融合ARNN和LSTM的森林火灾成灾面积预测方法的准确率分别提高了93.31%、47.74%、55.27%和9.39%。可见，所提方法能够高效地、较为精准地预测森林火灾成灾面积。

关键词: 自动储备池神经网络, 长短时记忆网络, 火灾成灾面积预测, 时序预测, 储备池计算