煤自燃是矿井火灾事故的重要成因，采空区是煤自燃的主要高发地点。结合日益复杂的矿井情况探究煤自燃的形成因素以及煤自燃的预测预警方法，能够有效降低采空区煤自燃事故发生可能性，而温度预测是最直接的监测预警方法之一。因此本文从煤自燃影响因素入手，采用大型煤自然发火实验台，最大限度模拟现场采空区煤自燃的温度演化以及气体分布，通过构建煤自燃温度动态预测模型，为采空区煤自燃的实际监测提供方法手段，对推进煤矿安全智能化发展具有重要意义。

采用1 kg煤自燃程序升温实验，得到不同煤样粒径、空气流量条件下的煤自燃过程中产生的碳氧化合物、碳氢化合物生成量以及生成速率等参数，均随着氧化温度的升高呈阶段性增加，并逐渐逼近指数型增长。根据煤自燃特征温度点，将煤自燃过程划分为平稳氧化、加速氧化以及突变氧化三个阶段，确定了CO、CH₄、C₂H₄、C₂H₆、第一火灾系数系数R₁、第二火灾系数R₂等能够表征煤自燃程度的多元气体指标。此外，采用2.0 t煤自然发火实验，通过计算时间尺度、空间尺度下的气体浓度变化以及全测点温度运移情况，可以得到煤自然发火过程中，由于空气流量的分布不均，在中心轴明显存在风量降低而升温速率逐渐减小的现象，煤自然发火实验炉内的温度演化与气体分布具备明显的“层级效应”，印证了煤自燃温度与气体产物分布之间存在非线性关联。

通过MATLAB R2022b软件的Neural Network Fitting框架，构建了煤自然发火实验中的温度动态预测模型，以进风口距离为切入点，将实际煤自燃过程中的采空区横向环境模拟为煤自然发火实验台纵向风流的三种层级情况，将模型分为进风口模型，一般层模型以及出风口模型，分别对应实际环境中较大空气流量、风量较小但煤的厚度较大、风量较小但存在气体聚集等环境条件。经测试，三种模型预测精度分别达到了0.85、0.92、0.96，MSE分别为89.5、35.23、16.30，基本符合要求，其中出风口模型通过GA优化算法改进后，RMSE调整为2.311，模型性能和泛化能力得到进一步提高，研究为煤矿防灭火的智能化发展提供了新思路。

Spontaneous coal combustion is an important cause of mine fire accidents, and goaf areas are the main places where spontaneous coal combustion occurs. Combining the increasingly complex mine conditions to explore the formation factors of coal spontaneous combustion and the prediction and early warning methods of coal spontaneous combustion can effectively reduce the possibility of coal spontaneous combustion accidents in goaf areas, and temperature prediction is one of the most direct monitoring and early warning methods. Therefore, this article starts from the influencing factors of coal spontaneous combustion and uses a large-scale coal spontaneous combustion experimental platform to simulate the temperature evolution and gas distribution of coal spontaneous combustion in the goaf area to the maximum extent. By constructing a dynamic prediction model of coal spontaneous combustion temperature, it provides the actual situation of coal spontaneous combustion in the goaf area. Monitoring provides methods and means, which is of great significance to promoting the safe and intelligent development of coal mines.

Using a 1 kg coal spontaneous combustion programmed temperature-raising experiment, parameters such as the amount of carbon oxides and hydrocarbons produced during the coal spontaneous combustion process under different coal sample particle sizes and air flow conditions were obtained, as well as the generation rate. It rises in stages and gradually approaches exponential growth. According to the characteristic temperature point of coal spontaneous combustion, the coal spontaneous combustion process is divided into three stages: stable oxidation, accelerated oxidation and sudden oxidation. It is determined that CO, CH₄, C₂H₄, C₂H₆, the first fire coefficient coefficient R₁, the second fire coefficient R₂, which can characterize Multi-element gas index of coal spontaneous combustion degree. In addition, a 2.0 t coal spontaneous ignition experiment was used to calculate the collected gas concentration changes at time scales and spatial scales and the temperature migration of all measuring points. It was found that during the coal spontaneous ignition process, due to the uneven distribution of air flow, there are obvious random phenomena. The coal spontaneous combustion reaction gradually decreases as the air volume decreases. The temperature evolution and gas distribution in the coal spontaneous combustion experimental furnace are divided into obvious "level effects", which confirms the nonlinear relationship between coal spontaneous combustion temperature and gas product distribution.

Through the Neural Network Fitting framework of MATLAB R2022b software, a temperature dynamic prediction model in the coal spontaneous combustion experiment was constructed. Taking the distance from the air inlet as the entry point, the lateral environment of the goaf during the actual coal spontaneous combustion process was simulated as a coal spontaneous combustion experimental platform. For the three levels of longitudinal air flow, the model is divided into air inlet model, general layer model and air outlet model, which respectively correspond to larger air flow in the actual environment, smaller air volume but thicker coal, and smaller air volume but presence of gas. gathering conditions. After testing, the model prediction accuracy reached 0.85, 0.92, and 0.96 respectively, and the MSE was 89.5, 35.23, and 16.30 respectively, which basically met the requirements. After the air outlet model was improved through the GA optimization algorithm, the RMSE was adjusted to 2.311, and the model performance and generalization ability were has been further improved, and the research provides new ideas for the intelligent development of coal mine fire prevention and extinguishing.