Beijing Tianma Intelligent Control Technology Co., Ltd.

针对现有工作面周期来压预测方法精度不足、泛化性较差和算力要求高等问题，提出了一种基于动态自适应旗鱼优化BP神经网络（DASFO−BP）的工作面周期来压预测模型。通过分析工作面周期来压机理，得到与来压相关的影响因素，通过皮尔逊相关系数确定对来压具有显著影响的因素（推进速度、直接顶厚度、基本顶厚度、采高、煤层倾角和倾向长度）作为预测模型输入，并以下次来压强度和来压步距作为预测模型输出。针对旗鱼优化（SFO）算法鲁棒性不足的问题，提出了动态自适应优化策略对SFO算法进行改进，即在优化前期利用SFO达到快速收敛的目的，中期则借助秃鹰搜索（BES）跳出局部最优，后期发挥粒子群优化（PSO）深度搜索的优势来提高解的精度。通过改进后的动态自适应旗鱼优化（DASFO）算法对BP神经网络的超参数进行训练，构建了基于DASFO−BP的来压预测模型。实验结果表明：DASFO算法在单峰和多峰测试函数上均能实现快速收敛；与BP，SFO−BP和NCPSO−BP相比，DASFO−BP对周期来压强度和步距的预测值与真实值更为接近，具有更高的精度，拟合能力和泛化能力强，能够准确预测下一周期来压分布情况。

In order to solve the problems of insufficient precision, poor generalization, and high computational requirements of existing methods for periodic pressure prediction of working face, a periodic pressure prediction model of working face based on dynamic adaptive sailfish optimization BP neural network (DASFO-BP) is proposed. By analyzing the mechanism of working face periodic pressure, the influencing factors related to pressure are obtained. The Pearson correlation coefficient is used to determine the factors that have a significant impact on pressure (advance speed, direct roof thickness, basic roof thickness, mining height, coal seam dip angle, and dip length) as inputs for the prediction model. The subsequent pressure intensity and pressure step distance are used as outputs for the prediction model. A dynamic adaptive optimization strategy is proposed to improve the robustness of the sailfish optimization (SFO) algorithm. In the early stage of optimization, SFO is used to achieve fast convergence, while in the middle stage, bald eagle search (BES) is used to escape local optima. In the later stage, the advantage of particle swarm optimization (PSO) deep search is utilized to improve the precision of the solution. A dynamic adaptive sailfish optimization (DASFO) algorithm is improved to train the hyperparameters of the BP neural network, and a pressure prediction model based on DASFO-BP is constructed. The experimental results indicate that the DASFO algorithm can achieve fast convergence on both unimodal and multimodal test functions. Compared with BP, SFO-BP, and NCPSO-BP, DASFO-BP has higher precision in predicting the intensity and step distance of periodic pressure, and has strong generalization ability and fitting capability. It can accurately predict the pressure and its distribution in the next period.

国家重点研发计划项目（2023YFC2907504）。

姚钰鹏，熊武. 基于动态自适应旗鱼优化BP神经网络的工作面周期来压预测[J]. 工矿自动化，2024，50（8）：30-37.

YAO Yupeng, XIONG Wu. Periodic pressure prediction of working face based on dynamic adaptive sailfish optimization BP neural network[J]. Journal of Mine Automation，2024，50（8）：30-37.