基于深度学习的采区构造煤分布动态预测模型研究

Coal-and-gas outburst is one of the main dynamic disasters in China and has positive correlation associations with deformation degree and thickness of tectonically deformed coal. Generally the type and thickness of tectonically deformed coal are identified by the logging curve and then the distribution of tectonically deformed coal is predicted using seismic attributes or inversion. Although the identifying accuracy of logging curve is higher than that of seismic data, well logs are sparse. The logging-constrained seismic prediction method cannot solve the distribution prediction of tectonically deformed coal. Relative to seismic and well logs, the type and thickness of tectonically deformed coal in the mining process are the direct data. It is updated dynamically with the mining. At present, there is a lack quantitative, dynamic distribution prediction method and technology to combine the indirect data such as seismic data and well logs with actual mining data. Deep learning is a kind of machine learning method, which can be used to represent the different abstract levels of the original data and improve the accuracy of prediction. Through the parallel processing of indirect data and real time update, the nonlinear relationship can be analyzed between the combined data and the distribution of tectonically deformed coal. The distribution dynamic prediction model of tectonically deformed coal can be established based on deep learning and it can improve forecast accuracy and reliability. Therefore, this prediction can help the coal producers to make a more specific and accurate prediction of coal-and-gas outburst.

煤与瓦斯突出是我国矿井主要动力灾害之一，与构造煤变形程度及厚度正相关。一般通过测井曲线识别构造煤类型和厚度，再利用地震属性或反演预测采区构造煤分布。虽然测井曲线的识别精度比地震数据高，但其较稀少，测井约束的地震预测方法还不能解决采区构造煤分布的预测问题。相对于地震和测井等间接数据来说，矿井采掘过程中实际揭露的构造煤类型和厚度是直接数据，并且随着采掘而动态更新。目前缺少一种能将地震数据和测井数据等间接数据和实际采掘数据这一直接数据有机结合，定量、动态预测采区构造煤分布的方法和技术。深度学习技术通过多层次学习，得到对原始数据的不同抽象层表示，进而提高预测准确性的一种机器学习方法。本次通过对间接数据和实时更新直接数据的并行处理，关联分析其与采区构造煤分布的非线性关系，建立基于深度学习的采区构造煤分布动态预测模型，提升预测的精度和可靠性，为煤与瓦斯突出动力灾害的有效防治，提供更可靠的预测成果。

煤与瓦斯突出是我国矿井主要动力灾害之一，与构造煤变形程度及厚度正相关。深度学习技术通过多层次学习，得到对原始数据的不同抽象层表示，进而提高预测准确性的一种机器学习方法。深度置信网络是深度学习的主流方法之一。为了预测煤层的构造煤厚度分布，我们提出了一种基于深度置信网络的构造煤分布预测方法。首先，我们通过遗传算法结合BP算法构建了构造煤分布预测相关敏感地震属性集；然后，建立了基于深度置信网络的采区构造煤分布预测模型，在含构造煤的地质概念模型数据和实际采区数据中分别验证了模型的可靠性；将真实的构造煤厚度与预测的构造煤厚度进行了比较，评价了不同模型的预测精度，即深度置信网络模型、支持向量机模型和极限学习机模型。最后，利用深度置信模型对芦岭煤矿8煤层构造煤分布进行了预测。研究表明，构造煤厚度的预测分布遵循构造煤发育的区域特征。同时，我们分析了影响构造煤分布的主要因素，构造煤分布与埋藏深度、煤层厚度和构造发育呈正相关。项目的研究为煤与瓦斯突出动力灾害的有效防治，提供更可靠的预测成果。