基于ISSA−LSTM的浓缩池溢流浓度预测

Prediction of overflow concentration of thickener based on ISSA-LSTM

浓缩池溢流浓度监测是实现煤泥水智能加药的关键。针对基于传感器的溢流浓度监测方式会导致絮凝剂调节滞后的问题,提出了一种基于改进麻雀搜索算法(ISSA)−长短期记忆(LSTM)的浓缩池溢流浓度预测方法。首先,对浓缩生产过程中的多参数时间序列进行相关性分析和预处理,得到输入变量。其次,采用多策略联合改进麻雀搜索算法(SSA):引入Tent混沌映射对麻雀种群进行初始化,以保证种群多样性,加快算法收敛速度;用螺旋捕食策略改进SSA的寻优过程,以兼顾局部开发和全局搜索能力;用萤火虫扰动策略对麻雀搜索结果进行扰动,以提高全局搜索能力,避免算法陷入局部最优。然后,采用ISSA优化双层LSTM网络模型的超参数。最后,构建基于ISSA−LSTM的浓缩池溢流浓度预测模型,进行在线监测。实验结果表明:①选取Ackley函数和Rastrigin函数作为测试函数,得出ISSA的全局寻优性能和收敛速度均优于粒子群优化(PSO)算法、鲸鱼优化算法(WOA)和标准SSA。②3种改进策略中,螺旋捕食策略对ISSA性能的提升起主导作用,混沌映射和萤火虫扰动策略协调算法的收敛速度和全局搜索能力,进一步提升算法寻优性能。③采用ISSA优化LSTM的超参数,解决了依靠主观经验取值时存在的欠拟合或过拟合问题,ISSA−LSTM模型的溢流浓度预测精度达97.26%,高于双层LSTM、SSA−LSTM、最小二乘支持向量机(LSSVM)等模型。④数据预处理可以提升模型的精度,降噪后溢流浓度预测精度比降噪前提升了30.25%。

The monitoring of the overflow concentration of the thickener is the key to realize intelligent dosing of coal slurry.The overflow concentration monitoring method based on the sensor will lead to the delay of flocculant regulation.In order to solve the above problem,a prediction method of overflow concentration of thickener based on improved sparrow search algorithm(ISSA)and long-short term memory(LSTM)is proposed.Firstly,the correlation analysis and pretreatment of multi-parameter time series in the process of concentration production are carried out to obtain the input variables.Secondly,the multi-strategies are combined to improve sparrow search algorithm(SSA).Tent chaotic map is introduced to initialize the sparrow population to ensure population diversity and speed up algorithm convergence.The optimization process of SSA is improved by using the spiral predation strategy to balance both local development and global search capabilities.The firefly perturbation strategy is used to perturb the sparrow search results to improve the global search performance and avoid the algorithm falling into local optimization.Thirdly,ISSA is used to optimize the hyperparameters of the two-layer LSTM network model.Finally,the overflow concentration prediction model based on ISSA-LSTM is established for on-line monitoring.The experimental results show the following points.①The Ackley function and Rastigin function are selected as test functions.It is concluded that ISSA's global optimization capability and convergence speed are better than those of the particle swarm optimization(PSO)algorithm,whale optimization algorithm(WOA)and standard SSA.②Among the three improved strategies,the spiral predation strategy plays a leading role in improving the performance of ISSA.The chaotic map and the firefly perturbation strategy coordinate the convergence speed and global search capability of the algorithm to further improve the optimization performance of the algorithm.③ISSA is used to optimize the hyperparameters of LSTM,which solves the problem of under-fitting or over-fitting when the values are determined by subjective experience.The prediction precision of overflow concentration of the ISSA-LSTM model reaches 97.26%,which is higher than that of double-layer LSTM,SSA-LSTM,and least square support vector machine(LSSVM)models.④Data pretreatment can improve the precision of the model,and the prediction precision of overflow concentration after noise reduction is improved by 30.25%compared with that before noise reduction.