煤自燃是煤矿开采过程中的主要自然灾害之一。我国大部分矿井为自燃和易自燃煤层，准确预测煤自燃危险性对于煤火灾害防治具有重大意义。但是，煤自燃的物理化学过程十分复杂，且影响因素众多，这给煤自燃危险性的预测带来很大的挑战。利用人工智能的理论与方法加强对煤自燃危险性预测技术的研究，对于提升煤矿安全生产智能化管控水平至关重要。本文选择了煤自然发火期、煤自燃交叉点温度（CPT）和煤自燃CO浓度作为煤自燃危险性评估指标，并基于实验和文献综述建立了数据集，结合机器学习理论与方法对煤自燃特征信息进行数据挖掘，构建了煤自燃危险性预测模型，为实现煤火灾害隐患和事故的超前预测提供了基础。

首先，建立了煤自然发火期、CPT和煤自燃CO浓度序列数据集并对3个数据集进行了预处理，包括数据缺失值处理、特征变量分布检验、特征变量相关性分析以及数据集归一化处理。从而降低了数据集噪点、变量依赖性以及实验系统性误差，提高了模型的泛化能力。

运用随机森林（RF）和多层感知机（MLP）2种算法分别建立了煤自然发火期预测模型和CPT预测模型，实现了对不同环境条件下不同煤种的发火期和CPT的预测，表征了复杂的物理化学反应和内外影响因素对煤自燃危险性的影响。在模型训练过程中，引入了网格搜索法、K折交叉验证、学习曲线以及3个模型性能检验指标，优化了模型的超参数和算法，评估检验了模型的状态和性能。结果表明基于RF算法和MLP算法的煤自然发火期预测模型均可利用煤自燃影响因素预测煤自然发火期，误差小于20%且 RF模型泛化能力更强；基于RF算法和MLP算法的CPT预测模型可以构建煤自燃内外影响因素与CPT之间的关系，两个模型的预测误差低于90%且RF模型泛化能力更强。

基于循环神经网络（RNN）、长短期记忆网络（LSTM）和门控循环单元（GRU）3种算法，建立了CO浓度动态序列预测模型，模型可准确预测未来时刻CO浓度变化趋势。在模型构建过程中，添加了dropout类避免模型出现过拟合，引入了均方误差（MSE）和3个模型性能检验指标，分析优化了CO预测模型的参数，检验了模型性能。研究结果表明，RNN、LSTM和GRU模型均能实现对CO浓度的动态预测且误差小于1%；另外，在同一序列数据下LSTM模型预测精度最高，其次是RNN模型和GRU模型。

Coal spontaneous combustion is one of the main natural disasters during coal mining. Most mines in China are coal seams that are prone to spontaneous combustion or spontaneous combustion. Accurately predicting the risk of coal spontaneous combustion is of great significance for the prevention and control of coal fire disasters. However, the physicochemical processes of coal spontaneous combustion are very complex and there are many influencing factors, which bring great challenges to the prediction of coal spontaneous combustion. Therefore, strengthening the research on coal spontaneous combustion risk prediction by using the theory and method of artificial intelligence is essential to improve the intelligent control of coal mine safety production. In this paper, coal spontaneous combustion period, coal spontaneous combustion crossing point temperature (CPT) and coal spontaneous combustion CO concentration were selected as the evaluation indicators for the coal spontaneous combustion risk and establishes the database through experiments and literature review. Combining theories and methods of machine learning to data mine coal spontaneous combustion features information, and coal spontaneous combustion risk prediction models are constructed. It provides basic to realize the advance prediction of coal fire disaster potential and accidents.

Firstly, the datasets of coal spontaneous combustion period, the datasets of CPT and sequential dataset of coal spontaneous combustion CO concentration were established based on experimental data and literature review. The three datasets were then preprocessed to improve the generalization capability of the models including missing value treatment, distribution test of the feature variables, correlation analysis of the feature variables, and normalization of the datasets to reduce the datasets noise, variable dependence, and experimental systematic errors.

Coal spontaneous combustion period prediction model and CPT prediction model are developed based on the coal spontaneous combustion period dataset and CPT dataset using both random forest (RF) and multilayer perceptron (MLP) algorithms. The prediction of the spontaneous combustion period and CPT of different coal types under different environmental conditions is realized, and the influence of complex physicochemical reactions and internal and external influences on the risk of coal spontaneous combustion is characterized. During the training process of two models, grid-search method, K-fold crossing validation, learning curve, and three model performance inspection indicators were introduced to optimize the hyperparameters, optimize the model algorithm, evaluate the model state, and test the model performance for the coal spontaneous combustion period model and CPT model. The results show that the coal spontaneous combustion period prediction model based on the RF algorithm and the MLP algorithm can use internal and external influences on coal spontaneous combustion to predict the coal spontaneous combustion period with an error of less than 20% and the RF model generalizes better; The CPT prediction model based on RF algorithm and MLP algorithm can build the relationship between the influencing factors of coal spontaneous combustion and CPTs, and the prediction error of the two models is less than 10% and the RF model generalizes better.

Based on the three algorithms of recurrent neural network (RNN), long short-term memory network (LSTM) and gated recurrent unit (GRU), a dynamic sequence prediction model of CO concentration was developed, and the model could accurately predict the trend of CO concentration changes in the future. And in the process of model construction, dropout class was added to avoid overfitting of the model, then mean square error (MSE) and three model performance inspection were introduced to analyze and optimize the parameters of CO models and test the prediction model performance. The results show that the RNN, LSTM and GRU models can achieve the dynamic prediction of CO concentration with the error less than 1%; in addition, the LSTM model has the highest prediction accuracy under the same sequence data, followed by the RNN model and GRU model.