基于改进极限学习机的水体溶解氧预测方法

为了有效地指导水产养殖生产，提高溶解氧浓度预测的精度，提出了基于因子筛选和改进极限学习机（Extreme Learning Machine，ELM）的水产养殖溶解氧预测模型。首先，利用皮尔森相关系数法计算各影响因子与溶解氧浓度间的相关系数，提取强关联因子，降低预测模型的输入量维度；采用偏最小二乘算法（Partial Least Square, PLS）优化传统ELM神经网络，避免网络中隐含层共线性问题，保障输出权值的稳定性；然后，结合新型激活函数，构建水体溶解氧浓度预测模型。最后，将SPLS-ELM（Selection Based Partial Least Square Optimized Extreme Learning Machine）预测模型应用到江苏省无锡市南泉基地某试验池塘的水体溶解氧预测中。试验结果表明：该模型的预测均方根误差为0.3232，与最小二乘支持向量机（Least Square Support Vector Machine，LSSVM）、BP神经网络、粒子群（Particle Swarm Optimization，PSO）优化LSSVM和遗传算法（Genetic Algorithm, GA）优化BP神经网络相比分别降低40.98%、44.48%、34.73%和44.18%。且该模型的运行时间仅0.6231s，预测精度和运行效率明显优于其他模型。该模型的溶解氧预测曲线接近真实溶解氧变化曲线，能够满足水产养殖实际生产对水体溶解氧预测的要求。

Dissolved oxygen prediction for water quality of aquaculture using improved ELM network

Highly accurate monitoring of water quality can efficiently provide scientific data to intensive aquaculture production. One of the most important parameters, dissolved oxygen (DO) content can be used to determine the fish survival rate in aquaculture water quality monitoring. However, the dissolved oxygen content can greatly vary in complex conditions, thereby to make it difficult to gain the high precision prediction. In this study, an improved extreme learning machine (ELM) neural network based on factor selection (SPLS-ELM) was proposed to forecast dissolved oxygen. First, Pearson correlation coefficient method was used to calculate the weights of other factors on dissolved oxygen. The strong correlation factors were extracted according to the obtained weights. The strong correlation factors were selected as the input data for a prediction model with reduced dimension. The key factors included water temperature, pH, temperature, humidity, illuminance, photosynthetically active radiation, irradiance and wind speed. Partial least-squares (PLS) was utilized to optimize the ELM neural network, in order to avoid high collinearity when the redundant data was input into traditional ELM, further to ensure the stability of output weight coefficients. Then, the dissolved oxygen prediction model SPLS-ELM was constructed based on the new activation function with good generalization. Finally, to verify the proposed SPLS-ELM prediction model, various experiments were performed on the monitoring of built-in water quality in Nanquan Aquaculture Base, Jiangsu Province, from July 1st, 2019 to July 30th, 2020. The prediction models were used to compare, including Least squares support vector machine (LSSVM), BP, particle swarm optimized LSSVM (PSO-LSSVM) and genetic algorithm optimized BP neural network (GA-BP) models. The experimental results showed that the error of root mean square (RMS) of SPLS-ELM was 0.3232, indicating the increase by 40.98%, 44.48%, 34.73% and 44.18%, compared with LSSVM, BP, PSO- LSSVM and GA-BP prediction model, respectively. The RMS error of SPLS-ELM improved by 27.24% and 46.82%, respectively, compared with PLS-ELM and ELM prediction model. The accuracy of the presented SPLS-ELM model was obviously higher than that of the counterpart models. The run time of SPLS-ELM prediction model was just 0.6231s. The efficiency of SPLS-ELM improved by about 3 times and 10 times, compared with than of LSSVM and BP, respectively. Meanwhile, the prediction curve of dissolved oxygen was closed to the real observed values. A better prediction performance was achieved by the improved operations of factor section, PLS algorithm and new activation function. The proposed SPLS-ELM can overcome the problem of collinearity in redundant input for the reliable prediction. SPLS-ELM can be expected to serves as the prediction of dissolved oxygen for water quality monitoring in real aquaculture.