CN103809557B

CN103809557B - A kind of sewage disposal process optimal control method based on neutral net

Info

Publication number: CN103809557B
Application number: CN201310745867.5A
Authority: CN
Inventors: 乔俊飞; 王莉莉; 韩红桂; 赵慢
Original assignee: Beijing University of Technology
Current assignee: Beijing University of Technology
Priority date: 2013-12-30
Filing date: 2013-12-30
Publication date: 2016-08-17
Anticipated expiration: 2033-12-30
Also published as: CN103809557A

Abstract

The sewage treatment process is a highly nonlinear, time-varying and complex process. It is very challenging to reduce the energy consumption of the operation under the premise of meeting the effluent water quality. Aiming at the problem of high energy consumption in the process of sewage treatment, the present invention proposes a neural network-based optimization control method, which mainly includes two neural networks, one of which is used to establish a prediction model for the process of sewage treatment to achieve performance Index prediction; Another neural network optimizes the set value of the control variable in real time according to the system state and the predicted performance index. Finally, the optimized control of dissolved oxygen concentration and nitrate nitrogen concentration can not only meet the requirements of effluent water quality, but also effectively reduce the operating cost of the sewage treatment process.

Description

An Optimal Control Method for Sewage Treatment Process Based on Neural Network

技术领域technical field

本发明针对污水处理过程能耗过高的问题，在BSM1中利用神经网络对污水处理过程中的溶解氧浓度和硝态氮浓度进行了优化控制。神经网络是智能控制技术的主要分支之一，基于神经网络的污水处理过程优化控制不但属于水处理领域，还属于智能优化控制领域。Aiming at the problem of high energy consumption in the sewage treatment process, the invention utilizes the neural network to optimize the control of the dissolved oxygen concentration and the nitrate nitrogen concentration in the sewage treatment process in the BSM1. Neural network is one of the main branches of intelligent control technology. The optimal control of sewage treatment process based on neural network belongs not only to the field of water treatment, but also to the field of intelligent optimal control.

背景技术Background technique

随着城市化、工业化的不断提速，我国水环境已遭到严重破坏并有继续恶化的趋势。污水排放不但严重影响着居民的日常生活，而且破坏了大自然的生态平衡。为了降低污水的排放量，全国各地已经纷纷建立了污水处理厂，然而污水处理过程电能消耗过大、运行成本居高，导致我国污水处理厂“建得起而养不起”，研究污水处理过程优化控制实现节能降耗意义重大，是未来污水处理行业必然的发展趋势。因此，本发明的研究成果具有广阔的应用前景。With the continuous acceleration of urbanization and industrialization, my country's water environment has been severely damaged and has a tendency to continue to deteriorate. Sewage discharge not only seriously affects the daily life of residents, but also destroys the ecological balance of nature. In order to reduce the discharge of sewage, sewage treatment plants have been established all over the country. However, the excessive power consumption and high operating costs in the sewage treatment process have led to the fact that my country's sewage treatment plants are "affordable to build but unaffordable to maintain". Research on sewage treatment process Optimizing control to achieve energy saving and consumption reduction is of great significance and is an inevitable development trend of the sewage treatment industry in the future. Therefore, the research results of the present invention have broad application prospects.

在污水处理过程中，主要的控制变量是第五分区的溶解氧浓度和第二分区的硝态氮浓度，溶解氧浓度和硝态氮浓度的高低不但影响硝化过程和反硝化过程的进行，而且对污水处理过程中的曝气能耗和泵送能耗起主要作用。因此，动态调整溶解氧和硝态氮的浓度的设定值对于改善出水水质，降低能耗是非常有必要的。In the sewage treatment process, the main control variables are the dissolved oxygen concentration in the fifth zone and the nitrate nitrogen concentration in the second zone. It plays a major role in the aeration energy consumption and pumping energy consumption in the sewage treatment process. Therefore, it is very necessary to dynamically adjust the set values of the concentration of dissolved oxygen and nitrate nitrogen to improve the quality of effluent water and reduce energy consumption.

污水处理过程是一个具有高度非线性、大滞后、大时变、多变量耦合等特点的复杂系统，需要完成出水水质达标同时降低能耗的多目标优化，部分学者对其进行了研究并取得了阶段性的成果。基于模型预测控制MPC的优化控制方法，虽然能在一定程度上降低能耗，但是MPC以系统机理模型为核心，通过确定的数学手段计算最优的控制量，而由于污水处理过程的复杂性，要建立其精确的数学模型存在一定难度，而模型的准确性对控制性能影响较大，部分关键水质参数的不可在线测量也会降低控制的性能。基于遗传算法GA的优化控制方法，其计算量大，全局收敛速度较慢，同时其优化结果的精度受编码长度的控制。基于粒子群PSO算法的优化控制方案，不需要复杂的交叉变异，算法简单、参数少、易于实现。无论是遗传算法还是粒子群算法，像这种智能搜索算法在优化过程中需要大量的样本，不能够对优化信号进行实时更新，比较适用于离线优化。The sewage treatment process is a complex system with the characteristics of high nonlinearity, large lag, large time variation, and multivariable coupling. It needs to complete the multi-objective optimization of effluent water quality and reduce energy consumption. Some scholars have studied it and obtained results. staged results. Although the optimal control method based on model predictive control (MPC) can reduce energy consumption to a certain extent, MPC takes the system mechanism model as the core and calculates the optimal control quantity through certain mathematical means. Due to the complexity of the sewage treatment process, It is difficult to establish an accurate mathematical model, and the accuracy of the model has a great influence on the control performance, and the inability to measure some key water quality parameters online will also reduce the control performance. The optimization control method based on the genetic algorithm GA has a large amount of calculation and a slow global convergence speed, and the precision of the optimization result is controlled by the length of the code. The optimal control scheme based on the particle swarm PSO algorithm does not require complex cross-mutation, and the algorithm is simple, with few parameters and easy to implement. Whether it is a genetic algorithm or a particle swarm algorithm, such an intelligent search algorithm requires a large number of samples in the optimization process, and cannot update the optimization signal in real time, so it is more suitable for offline optimization.

神经网络具有非常强大的学习能力和自适应特性，能够对非线性系统进行高精度逼近。本发明提出了一种基于神经网络的污水处理过程优化控制方法，不但能否满足出水水质的要求，还能降低污水处理过程的能耗。The neural network has a very powerful learning ability and adaptive characteristics, and can approximate nonlinear systems with high precision. The invention proposes a neural network-based optimization control method for sewage treatment process, which can not only meet the requirements of effluent water quality, but also reduce energy consumption in the sewage treatment process.

发明内容Contents of the invention

基于神经网络的污水处理过程优化控制方法，主要包括三个部分：性能指标预测模型、神经网络优化层以及底层的控制器。利用性能指标预测模型预测未来时刻的性能指标函数值，根据当前的环境状态及上一时刻的控制变量的设定值，通过神经网络优化层使得性能指标预测模型预测出的性能指标函数达到最小，从而产生新的设定值传送到底层控制器，完成污水处理过程的优化控制。The optimization control method of sewage treatment process based on neural network mainly includes three parts: performance index prediction model, neural network optimization layer and the underlying controller. Use the performance index prediction model to predict the performance index function value in the future, according to the current environmental state and the set value of the control variable at the previous moment, through the neural network optimization layer, the performance index function predicted by the performance index prediction model is minimized, Thus, a new set value is generated and sent to the bottom controller to complete the optimal control of the sewage treatment process.

本发明采用了如下的技术方案及实现步骤：The present invention adopts following technical scheme and implementation steps:

基于神经网络的污水处理过程优化控制方法，其特征在于，包括以下步骤：The optimization control method of sewage treatment process based on neural network is characterized in that it comprises the following steps:

1.性能指标的建立1. Establishment of performance indicators

污水处理过程是一个具有多变量耦合、干扰严重、大滞后特性的高度非线性系统，而神经网络能够以其非常强大的非线性逼近能力，对复杂的控制系统进行建模。污水处理过程的优化问题需要综合考虑出水水质、曝气能耗以及泵送能耗等目标，是一个多目标优化问题。因此，定义优化的性能指标函数为The sewage treatment process is a highly nonlinear system with multi-variable coupling, serious interference, and large hysteresis characteristics, and the neural network can model the complex control system with its very powerful nonlinear approximation ability. The optimization problem of the sewage treatment process needs to comprehensively consider the goals of effluent water quality, aeration energy consumption and pumping energy consumption, which is a multi-objective optimization problem. Therefore, the optimized performance index function is defined as

J(k)＝α₁E(k)+α₂E_Q(k) (1)J(k)＝α ₁ E(k)+α ₂ E _Q (k) (1)

α₁、α₂为衡量能耗E和出水水质E_Q的权重因子，α₁、α₂∈[0,1]，并且α₁+α₂＝1，E为一个优化周期内的总能耗，是曝气能耗E_A和泵送能耗E_P之和，E_Q表示一个优化周期向受纳水体排放污染物需要支付的罚款。E、E_A、E_P、E_Q的表达式分别如(2)、(3)、(4)、(5)所示。α ₁ , α ₂ are weighting factors to measure energy consumption E and effluent water quality E _Q , α ₁ , α ₂ ∈ [0,1], and α ₁ + α ₂ = 1, E is the total energy consumption in an optimization cycle , is the sum of aeration energy consumption E _A and pumping energy consumption E _P , and E _Q represents the penalty that needs to be paid for discharging pollutants to the receiving water body in an optimization cycle. The expressions of E, E _A , E _P , and E _Q are shown in (2), (3), (4), and (5) respectively.

E(k)＝E_A(k)+E_P(k) (2)E(k)＝E _A (k)+E _P (k) (2)

${E E.}_{A A} ((k k)) = = \frac{{S S}_{o o,, s the s a a t t}}{T T \cdot &Center Dot; 1.8 1.8 \cdot &Center Dot; 10001000} {&Integral; &Integral;}_{k k T T}^{((k k + + 11)) T T} {Σ Σ}_{i i = = 11}^{i i = = 55} {V V}_{i i} \cdot &Center Dot; {K K}_{L L} {a a}_{i i} ((t t)) d d t t - - - - - - ((33))$

${E E.}_{P P} ((k k)) = = \frac{11}{T T} {&Integral; &Integral;}_{k k T T}^{((k k + + 11)) T T} 0.004 0.004 {Q Q}_{a a} ((t t)) + + 0.008 0.008 {Q Q}_{r r} ((t t)) + + 0.05 0.05 {Q Q}_{w w} ((t t)) d d t t - - - - - - ((44))$

${E E.}_{Q Q} ((k k)) = = \frac{11}{T T \cdot &Center Dot; 10001000} {&Integral; &Integral;}_{k k T T}^{((k k + + 11)) T T} (({B B}_{S S S S} \cdot \cdot {SS SS}_{e e} ((t t)) + + {B B}_{C C O o D D.} \cdot \cdot {COD COD}_{e e} ((t t)) + + {B B}_{N N O o} \cdot \cdot {S S}_{N N O o,, e e} ((t t)) + + {B B}_{N N k k j j,, e e} \cdot \cdot {S S}_{N N k k j j . . e e} ((t t)) + + {B B}_{B B O o D D. 55} \cdot \cdot {BOD BOD}_{e e} ((t t)))) d d t t - - - - - - ((55))$

k为时刻，T为优化周期，S_o,sat为溶解氧的饱和浓度，S_O,sat＝8mg/L，V_i、K_La_i(i＝1,2,3,4,5)分别为第一个到第五个反应池的体积和氧气传递系数，其中，V₁＝V₂＝1000m³，V₃＝V₄＝V₅＝1333m³，K_La₁＝K_La₂＝0，K_La₃＝K_La₄＝240d^-1，K_La₅和Q_a分别为控制第五个反应池中溶解氧浓度和第二个反应池中硝态氮浓度的操作变量，为控制器的输出，是两个变量，K_La₅变化范围为0～240d^-1，Q_a变化范围为0～5Q₀，Q₀为进水流量，Q_r为污泥回流量，Q_w为污泥排放量，B_SS、B_COD、B_NO、B_NKj、B_BOD5为出水SS_e、COD_e、S_NO,e、S_NKj,e、BOD_e对出水水质E_Q影响的权重因子，B_SS＝2，B_COD＝1，B_NO＝10，B_NKj＝30，B_BOD5＝2。k is the moment, T is the optimization period, S _o,sat is the saturation concentration of dissolved oxygen, S _O,sat =8mg/L, V _i , K _L a _i (i=1,2,3,4,5) respectively is the volume and oxygen transfer coefficient of the first to fifth reaction tanks, where, V ₁ =V ₂ =1000m ³ , V ₃ =V ₄ =V ₅ =1333m ³ , K _L a ₁ =K _L a ₂ = 0, K _L a ₃ ＝K _L a ₄ ＝240d ^-1 , K _L a ₅ and Q _a are operating variables for controlling the dissolved oxygen concentration in the fifth reaction tank and the nitrate nitrogen concentration in the second reaction tank respectively, The output of the controller is two variables, K _L a ₅ ranges from 0 to 240d ^-1 , Q _a ranges from 0 to 5Q ₀ , Q ₀ is the influent flow, Q _r is the sludge return flow, Q _w is the sludge discharge, B _SS , B _COD , B _NO , B _NKj , B _BOD5 are the weight factors of the influence of effluent SS _e , COD _e , S _NO,e , S _NKj,e , and BOD _e on the effluent quality E _Q , B _SS =2, B _COD =1, B _NO =10, B _NKj =30, B _BOD5 =2.

2.k时刻，溶解氧浓度和硝态氮浓度的预测优化控制2. Predictive optimization control of dissolved oxygen concentration and nitrate nitrogen concentration at time k

2.1首先建立性能指标的预测模型，预测k时刻的性能指标值J'(k)，采用回声状态网络ESN构建预测模型，预测ESN的输入层、输出层、动态储备池DR分别包含K、L、N个神经元，W_P ⁱⁿ为预测ESN输入层到内部状态储备池的连接权值矩阵，W_P为预测ESN动态储备池内部的权值矩阵，其神经元之间为稀疏连接，定义稀疏度SD为动态储备池内相互连接的神经元个数与总神经元个数之比，SD一般取值2％～5％，以保证动态储备池丰富动态特性，W_P ^back为预测ESN输出层到内部状态储备池的连接权值矩阵，W_P ⁱⁿ、W_P、W_P ^back维数分别为N×K、N×N、N×L，初始值均为随机产生，一经产生，就不再变化，预测ESN的权谱半径小于1，以保证网络的稳定性。预测ESN的输入是k时刻溶解氧浓度r₁(k)和硝态氮浓度r₂(k)的设定值，输出是预测的性能指标值，预测ESN的输入为：2.1 Firstly, establish the prediction model of the performance index, predict the performance index value J'(k) at time k, use the echo state network ESN to construct the prediction model, and predict that the input layer, output layer, and dynamic reserve pool DR of the ESN include K, L, N neurons, W _P ⁱⁿ is to predict the connection weight matrix from the input layer of ESN to the internal state reserve pool, W _P is to predict the weight matrix inside the dynamic reserve pool of ESN, the neurons are sparsely connected, and the sparsity is defined SD is the ratio of the number of interconnected neurons to the total number of neurons in the dynamic reserve pool. SD generally takes a value of 2% to 5% to ensure the rich dynamic characteristics of the dynamic reserve pool. W _P ^back is the prediction of the ESN output layer to the internal The connection weight matrix of the state reserve pool, the dimensions of W _P ⁱⁿ , W _P , and W _P ^back are N×K, N×N, and N×L respectively, and the initial values are randomly generated. Once generated, they will not change again. It is predicted that the weight spectrum radius of ESN is less than 1 to ensure the stability of the network. The input of predicting ESN is the set value of dissolved oxygen concentration r ₁ (k) and nitrate nitrogen concentration r ₂ (k) at time k, and the output is the predicted performance index value. The input of predicting ESN is:

R(k)＝[r₁(k) r₂(k)]^T (6)R(k)＝[r ₁ (k) r ₂ (k)] ^T (6)

动态储备池的输出为：The output of the dynamic reserve pool is:

${x x}_{P P} ((k k)) = = f f (({W W}_{P P}^{i i n no} R R ((k k)) + + {W W}_{P P} {x x}_{P P} ((k k - - 11)) + + {W W}_{P P}^{b b a a c c k k} {J J}^{' '} ((k k - - 11)))) - - - - - - ((77))$

J'(k-1)为上一时刻预测ESN的输出，f为内部神经元激活函数，取为Sigmoid函数。J'(k-1) is the output of predicted ESN at the previous moment, and f is the internal neuron activation function, which is taken as the Sigmoid function.

ESN的输出为性能指标函数：The output of ESN is the performance index function:

${J J}^{' '} ((k k)) = = {f f}^{o o u u t t} (({W W}_{P P}^{o o u u t t} {x x}_{P P} ((k k)))) - - - - - - ((88))$

式中，W_P ^out为预测ESN内部状态储备池到输出的连接权值矩阵，维数为L×N，是网络训练过程中唯一需要进行调整的权值，f^out为输出层传递函数，取为线性函数，输出J'(k)为k时刻的性能指标函数值。In the formula, W _P ^out is the connection weight matrix for predicting the ESN internal state reserve pool to the output, the dimension is L×N, which is the only weight that needs to be adjusted during the network training process, and f ^out is the transfer function of the output layer, which is taken as is a linear function, and the output J'(k) is the performance index function value at time k.

2.2将k时刻溶解氧浓度r₁(k)和硝态氮浓度r₂(k)的设定值作为污水处理过程的输入，得到在k时刻的实际性能指标函数值J(k)。修正预测模型网络权值的指标函数为：2.2 Take the set values of dissolved oxygen concentration r ₁ (k) and nitrate nitrogen concentration r ₂ (k) at time k as the input of the sewage treatment process to obtain the actual performance index function value J(k) at time k. The index function for modifying the network weights of the prediction model is:

$g g ((k k)) = = \frac{11}{22} {[[{J J}^{' '} ((k k)) - - J J ((k k))]]}^{22} - - - - - - ((99))$

权值调整公示为：The weight adjustment is announced as:

${ΔW ΔW}_{P P}^{o o u u t t} ((k k)) = = - - {η η}_{P P} ((k k)) {((\frac{\partial \partial g g ((k k))}{\partial \partial {W W}_{P P}^{o o u u t t} ((k k))}))}^{T T} - - - - - - ((1010))$

其中，η_P为学习速率。Among them, η _P is the learning rate.

2.3优化神经网络同样采用ESN，优化的输入层、输出层、动态储备池DR分别包含K′、L′、N′个神经元，W_o ⁱⁿ为优化ESN输入层到内部状态储备池的连接权值矩阵，W_o为优化ESN动态储备池内部的权值矩阵，其神经元之间为稀疏连接，稀疏度同样保持2％～5％，保证网络具有丰富的动态特性，W_o ^back为优化ESN输出层到内部状态储备池的连接权值矩阵，W_o ⁱⁿ、W_o、W_o ^back维数分别为N′×K′、N′×N′、N′×L′，初始值均为随机产生，一经产生，就不再变化，优化ESN的权谱半径同样应小于1，以保证网络的稳定性，f取为Sigmoid函数。2.3 Optimizing the neural network also uses ESN. The optimized input layer, output layer, and dynamic reserve pool DR respectively contain K′, L′, and N′ neurons, and W _o ⁱⁿ is the connection weight of the optimized ESN input layer to the internal state reserve pool. Value matrix, W _o is to optimize the weight matrix inside the ESN dynamic reserve pool, the neurons are sparsely connected, and the sparsity is also maintained at 2% to 5%, ensuring that the network has rich dynamic characteristics, and W _o ^back is to optimize ESN The connection weight matrix from the output layer to the internal state reserve pool, the dimensions of W _o ⁱⁿ , W _o , and W _o ^back are N′×K′, N′×N′, N′×L′ respectively, and the initial values are all random Once generated, it will no longer change. The weight spectrum radius of the optimized ESN should also be less than 1 to ensure the stability of the network, and f is taken as the Sigmoid function.

其输入为Its input is

h(k+1)＝[R^T(k-1),S(k)]^T (11)h(k+1)=[RT(k-1),S(k)] ^T ⁽ 11)

R^T(k-1)为k-1时刻溶解氧浓度和硝态氮浓度的设定值，S(k)为k时刻的环境状态，可表示为：R ^T (k-1) is the set value of dissolved oxygen concentration and nitrate nitrogen concentration at time k-1, and S(k) is the environmental state at time k, which can be expressed as:

S(k)＝[S_O(k),S_NO(k),S_NH(k),S_ND(k),X_ND(k),Q₀(k)] (12)S(k)＝[S _O (k), S _NO (k), S _NH (k), S _ND (k), X _ND (k), Q ₀ (k)] (12)

S_O(k)为k时刻进水溶解氧浓度、S_NO(k)为k时刻进水硝态氮浓度、S_NH(k)为k时刻进水氨浓度、S_ND(k)为k时刻进水可溶性有机氮、X_ND(k)为k时刻进水不可溶性有机氮、Q₀(k)为k时刻进水流量。S _O (k) is the influent dissolved oxygen concentration at time k, S _NO (k) is the influent nitrate nitrogen concentration at k time, S _NH (k) is the influent ammonia concentration at k time, S _ND (k) is the Influent soluble organic nitrogen, X _ND (k) is insoluble organic nitrogen at time k, Q ₀ (k) is influent flow at time k.

优化ESN的动态储备池输出为：The dynamic reserve pool output of optimized ESN is:

${x x}_{o o} ((k k + + 11)) = = f f (({W W}_{o o}^{i i n no} h h ((k k + + 11)) + + {W W}_{o o} {x x}_{o o} ((k k)) + + {W W}_{o o}^{b b a a c c k k} R R ((k k)))) - - - - - - ((1313))$

ESN的输出为设定值：The output of the ESN is the setpoint:

$R R ((k k + + 11)) = = {f f}^{o o u u t t} (({W W}_{o o}^{o o u u t t} {x x}_{o o} ((k k + + 11)))) - - - - - - ((1414))$

W_o ^out为优化ESN内部状态储备池到输出层的连接权值矩阵，维数为L′×N′，在学习过程中需要进行调整，f^out取为线性函数。W _o ^out is to optimize the connection weight matrix from the ESN internal state reserve pool to the output layer, the dimension is L′×N′, which needs to be adjusted during the learning process, and f ^out is taken as a linear function.

2.4将R(k+1)输入到性能指标预测模型中，得到预测ESN新的动态储备池输出为：2.4 Input R(k+1) into the performance index prediction model, and the output of the new dynamic reserve pool for predicting ESN is:

${x x}_{P P} ((k k + + 11)) = = f f (({W W}_{P P}^{i i n no} R R ((k k + + 11)) + + {W W}_{P P} {x x}_{P P} ((k k)) + + {W W}_{P P}^{b b a a c c k k} {J J}^{' '} ((k k)))) - - - - - - ((1515))$

则k+1时的性能指标预测值为：Then the predicted value of the performance index at k+1 is:

${J J}^{' '} ((k k + + 11)) = = {f f}^{o o u u t t} (({W W}_{P P}^{o o u u t t} {x x}_{P P} ((k k + + 11)))) - - - - - - ((1616))$

2.5对优化ESN进行权值调整公式如下：2.5 The weight adjustment formula for the optimized ESN is as follows:

${ΔW ΔW}_{o o}^{o o u u t t} ((k k + + 11)) = = - - η η ((k k)) {((\frac{\partial \partial {J J}^{' '} ((k k + + 11))}{\partial \partial {W W}_{o o}^{o o u u t t} ((k k + + 11))}))}^{T T} - - - - - - ((1717))$

2.6将优化后新产生的设定值作为k+1时刻的值输入到PID控制器中，得到污水处理过程k+1时刻实际输出J(k+1)，对性能指标预测模型权值重新调整，权值调整的指标函数为：2.6 Input the newly generated set value after optimization into the PID controller as the value at time k+1 to obtain the actual output J(k+1) of the sewage treatment process at time k+1, and readjust the weight of the performance index prediction model , the indicator function for weight adjustment is:

$g g ((k k + + 11)) = = \frac{11}{22} {[[{J J}^{' '} ((k k + + 11)) - - J J ((k k + + 11))]]}^{22} - - - - - - ((1818))$

3.将步骤2.1至步骤2.6中的k加n，变为k+n，重复步骤2.1至步骤2.6，n＝1,2,3,4…，进行循环，直到不再有进水数据。3. Add n to k in step 2.1 to step 2.6 to become k+n, repeat step 2.1 to step 2.6, n=1, 2, 3, 4..., and cycle until there is no more water inflow data.

本发明的创造性主要体现在：The inventiveness of the present invention is mainly reflected in:

本发明设计了污水处理过程上层的优化方法，该方法能够根据环境状态实时优化控制变量的设定值。其一，利用ESN建立了设定值R(k)与性能指标J(k)间的关系，即性能指标智能预测模型，通过预测模型能够得到未来时刻的性能指标值；其二，同样利用ESN对预测的性能指标进行优化，使其达到最小，进而获得性能指标值最小时的控制变量设定值。以上两部分够成的预测优化方法，属于本发明的保护范围。The invention designs an optimization method for the upper layer of the sewage treatment process, which can optimize the set value of the control variable in real time according to the environment state. First, the relationship between the set value R(k) and the performance index J(k) is established by using ESN, that is, the intelligent prediction model of the performance index, and the performance index value at the future time can be obtained through the prediction model; Optimize the predicted performance index to make it the minimum, and then obtain the set value of the control variable when the performance index value is the minimum. The prediction and optimization method completed by the above two parts belongs to the protection scope of the present invention.

本发明提出的基于ESN的污水处理过程优化控制方法，解决了模型预测控制中建立的机理模型不精确问题，克服了智能优化算法计算复杂度大、不能实时更新优化控制信号的缺点。The ESN-based sewage treatment process optimization control method proposed by the present invention solves the problem of inaccurate mechanism models established in model predictive control, and overcomes the shortcomings of intelligent optimization algorithms with high computational complexity and inability to update optimization control signals in real time.

附图说明Description of drawings

图1.污水处理过程基准模型Figure 1. Benchmark model of wastewater treatment process

图2.ESN网络拓扑结构图Figure 2. ESN network topology diagram

图3.预测优化控制结构图Figure 3. Predictive optimization control structure diagram

图4.溶解氧优化效果Figure 4. Dissolved oxygen optimization effect

图5.硝态氮优化效果Figure 5. Optimization effect of nitrate nitrogen

图6.出水BOD的效果比较Figure 6. Effect comparison of effluent BOD

图7.出水COD的效果比较Figure 7. Effect comparison of effluent COD

图8.出水TSS的效果比较Figure 8. Comparison of the effect of TSS in effluent

图9.出水氨氮的效果比较Figure 9. Effect comparison of effluent ammonia nitrogen

图10.出水总氮的效果比较Figure 10. Effect comparison of effluent total nitrogen

具体实施方式detailed description

文中的实验是基于BSM1模型晴朗天气下的数据进行的，具体步骤如下：The experiment in this paper is based on the data of the BSM1 model under sunny weather. The specific steps are as follows:

1.建立性能指标预测模型1. Establish performance index prediction model

在建立的性能指标中，α₁、α₂分别取为0.8、0.2，预测模型的输入是溶解氧浓度和硝态氮浓度的设定值，输出为性能指标值，内部神经元个数为45个，即预测模型的结构为2-45-1，初始化网络的权值，输入到内部状态的权值W_P ⁱⁿ的维数为45×2，内部状态之间的连接权值W_P的维数为45×45，内部状态到输出的权值W_P ^out的维数为1×45，输出到内部状态的权值W_P ^back的维数为45×1，稀疏度SD为5％，权谱半径为0.48。In the established performance index, α ₁ and α ₂ are taken as 0.8 and 0.2 respectively, the input of the prediction model is the set value of dissolved oxygen concentration and nitrate nitrogen concentration, the output is the performance index value, and the number of internal neurons is 45 One, that is, the structure of the prediction model is 2-45-1, the weight of the network is initialized, the dimension of the weight W _P ⁱⁿ input to the internal state is 45×2, and the dimension of the connection weight W _P between the internal states The number is 45×45, the dimension of the weight value W _P ^out from the internal state to the output is 1×45, the dimension of the weight value W _P ^back from the output to the internal state is 45×1, the sparsity SD is 5%, and the weight The spectral radius is 0.48.

2.神经网络优化层的建立2. Establishment of neural network optimization layer

优化神经网络的输入为上一时刻的设定值以及当前时刻的状态，包括：进水溶解氧浓度S_O、进水硝态氮浓度S_NO、进水氨浓度S_NH、进水可溶性有机氮S_ND、进水不可溶性有机氮X_ND和进水流量Q₀，输出为设定值，内部神经元个数同样为45，优化神经网络的结构为8-45-2，初始化网络的权值，输入到内部状态的权值W_o ⁱⁿ的维数为45×8，内部状态之间的连接权值W_o的维数为45×45，内部状态到输出的权值W_o ^out的维数为2×45，输出到内部状态的权值W_o ^back的维数为45×2，稀疏度SD为5％，权谱半径为0.57。The input of the optimized neural network is the set value at the previous moment and the state at the current moment, including: influent dissolved oxygen concentration S _O , influent nitrate nitrogen concentration S _NO , influent ammonia concentration S _NH , influent soluble organic nitrogen S _ND , water insoluble organic nitrogen X _ND and water flow Q ₀ , the output is the set value, the number of internal neurons is also 45, the structure of the optimized neural network is 8-45-2, and the weight of the network is initialized , the dimension of the weight W _o ⁱⁿ input to the internal state is 45×8, the dimension of the connection weight W _o between the internal states is 45×45, and the dimension of the weight W _o ^out from the internal state to the output is 2×45, the dimension of the weight W _o ^back output to the internal state is 45×2, the sparsity SD is 5%, and the weight spectrum radius is 0.57.

3.底层控制器3. Bottom Controller

底层控制为PID控制器，其参数分别取为：KP_O＝200，KI_O＝150，KD_O＝5，KP_NO＝80000，KI_NO＝7000，KD_NO＝400。The underlying control is a PID controller, and its parameters are respectively taken as: KP _O =200, _{KIO =150, KD O} ₌ 5, KP _NO =80000, KI _NO =7000, KD _NO =400.

4.通过迭代优化，可以得到溶解氧浓度和硝态氮浓度的优化结果如图3和图4中的红色线所示，可以看出，控制变量的设定值随着入水流量和入水组分(即系统的状态)实时地变化，蓝色的线为控制变量的PID控制效果。将神经网络优化控制与BSM1中缺省的PID闭环控制(溶解氧浓度和硝态氮的设定值为固定值，分别为2mg/L和1mg/L)相比较，出水BOD、出水COD、出水TSS、出水氨氮以及出水总氮的对比分别如图6、7、8、9、10所示，由图可知，出水BOD、出水COD及出水TSS在两种控制情况下前后变化不大，而随着对溶解氧浓度和硝态氮浓度的实时优化，溶解氧浓度的平均设定值与固定设定值2mg/L相比降低了，硝态氮浓度的平均设定值与固定设定值1mg/L相比升高了。优化控制与PID闭环控制相比，虽然EQ增大了0.881％，但是出水水质仍然满足国家排放标准，优化控制比PID闭环控制的泵送能耗PE增加了12.77％，曝气能耗AE减少了4.958％，致使总能耗(AE与PE之和)减少3.906％，达到了较好的节能降耗效果。4. Through iterative optimization, the optimization results of dissolved oxygen concentration and nitrate nitrogen concentration can be obtained, as shown in the red lines in Figure 3 and Figure 4. It can be seen that the set value of the control variable varies with the influent flow rate and influent water composition. (that is, the state of the system) changes in real time, and the blue line is the PID control effect of the control variable. Comparing the neural network optimal control with the default PID closed-loop control in BSM1 (the set values of dissolved oxygen concentration and nitrate nitrogen are fixed values, respectively 2mg/L and 1mg/L), the effluent BOD, effluent COD, effluent The comparisons of TSS, effluent ammonia nitrogen, and effluent total nitrogen are shown in Figures 6, 7, 8, 9, and 10, respectively. It can be seen from the figure that the effluent BOD, effluent COD, and effluent TSS did not change much under the two control conditions. With the real-time optimization of dissolved oxygen concentration and nitrate nitrogen concentration, the average set value of dissolved oxygen concentration is reduced compared with the fixed set value of 2mg/L, and the average set value of nitrate nitrogen concentration is compared with the fixed set value of 1mg /L compared to increased. Compared with the PID closed-loop control, the optimal control has increased the EQ by 0.881%, but the effluent water quality still meets the national discharge standards. Compared with the PID closed-loop control, the pumping energy consumption PE of the optimal control has increased by 12.77%, and the aeration energy consumption AE has decreased. 4.958%, resulting in a reduction of 3.906% in total energy consumption (the sum of AE and PE), achieving a better effect of energy saving and consumption reduction.

Claims

1. a sewage disposal process optimal control method based on neutral net, it is characterised in that comprise the following steps:

1) foundation of performance indications forecast model

The performance index function that definition optimizes is

J (k)=α₁E(k)+α₂E_Q(k) (1)

α₁、α₂For weighing energy consumption E and effluent quality E_QWeight factor, α₁、α₂∈ [0,1], and α₁+α₂=1, E are an optimization Total energy consumption in cycle, is aeration energy consumption E_AWith pumping energy consumption E_PSum；E、E_A、E_P、E_QExpression formula respectively as (2), (3), (4), shown in (5)；

E (k)=E_A(k)+E_P(k) (2)

E_{A} (k) = \frac{S_{o, s a t}}{T \cdot 1.8 \cdot 1000} {&Integral;}_{k T}^{(k + 1) T} Σ_{i = 1}^{i = 5} V_{i} \cdot K_{L} a_{i} (t) d t - - - (3)

E_{P} (k) = \frac{1}{T} {&Integral;}_{k T}^{(k + 1) T} 0.004 Q_{a} (t) + 0.008 Q_{r} (t) + 0.05 Q_{w} (t) d t - - - (4)

E_{Q} (k) = \frac{1}{T \cdot 1000} {&Integral;}_{k T}^{(k + 1) T} (B_{S S} \cdot {SS}_{e} (t) + B_{C O D} \cdot {COD}_{e} (t) + B_{N O} \cdot S_{N O, e} (t) + B_{N k j} \cdot S_{N k j, e} (t) + B_{B O D 5} \cdot {BOD}_{e} (t)) d t - - - (5)

K is the moment, and T is optimization cycle, S_o,satFor the saturated concentration of dissolved oxygen, S_o,sat=8mg/L, V_i、K_La_i(i=1,2,3,4, 5) it is respectively first volume to the 5th reaction tank and oxygen carry-over factor, wherein, V₁=V₂=1000m³, V₃=V₄=V₅ =1333m³, K_La₁=K_La₂=0, K_La₃=K_La₄=240d^-1, K_La₅And Q_aIt is respectively and controls oxygen transmission in the 5th reaction tank The performance variable of nitrate in coefficient and second reaction tank, for the output of controller, is two variablees, K_La₅Change model Enclose is 0～240d^-1, Q_aExcursion is 0～5Q₀, Q₀For flow of inlet water, Q_rFor sludge reflux amount, Q_wFor sludge discharge, B_SS、 B_COD、B_NO、B_NKj、B_BOD5For water outlet SS_e、COD_e、S_NO,e、S_NKj,e、BOD_eTo effluent quality E_QThe weight factor of impact, B_SS=2, B_COD=1, B_NO=10, B_NKj=30, B_BOD5=2；

2) prediction optimization of k moment, dissolved oxygen concentration and nitrate controls

Step 2.1 initially sets up the forecast model of performance indications, it was predicted that the performance index value J'(k in k moment), use echo state Network ESN builds forecast model, it was predicted that the input layer of ESN, output layer, dynamic reserve pool DR comprise K, L, N number of neuron respectively, W_P ⁱⁿFor the connection weight value matrix of prediction ESN input layer to internal state reserve pool, W_PWithin the prediction dynamic reserve pool of ESN Weight matrix, is partially connected between its neuron, and definition degree of rarefication SD is interconnective neuron in dynamic reserve pool Counting the ratio with total neuron number, SD value 2%～5%, to ensure that dynamic reserve pool enriches dynamic characteristic, W_P ^backFor prediction ESN output layer is to the connection weight value matrix of internal state reserve pool, W_P ⁱⁿ、W_P、W_P ^backDimension is respectively N × K, N × N, N × L, Initial value is and randomly generates, and once generation, the most no longer changes, it was predicted that the power spectral radius of ESN is less than 1, to ensure the steady of network Qualitative；The input of prediction ESN is k moment dissolved oxygen concentration r₁(k) and nitrate r₂K the setting value of (), output is prediction Performance index value, it was predicted that the input of ESN is:

R (k)=[r₁(k) r₂(k)]^T (6)

Dynamically reserve pool is output as:

x_{P} (k) = f (W_{P}^{i n} R (k) + W_{P} x_{P} (k - 1) + W_{P}^{b a c k} J^{'} (k - 1)) - - - (7)

J'(k-1) being a output upper moment predicting ESN, f is intrinsic nerve unit activation primitive, is taken as Sigmoid function；

ESN is output as performance index function:

J^{'} (k) = f^{o u t} (W_{P}^{o u t} x_{P} (k)) - - - (8)

In formula, W_P ^outFor the connection weight value matrix of prediction ESN internal state reserve pool to output, dimension is L × N, is network training During uniquely need the weights that are adjusted, f^outTransmit function for output layer, be taken as linear function, export J'(k) when being k The performance index function value carved；

Step 2.2 is by k moment dissolved oxygen concentration r₁(k) and nitrate r₂Defeated as sewage disposal process of the setting value of (k) Enter, obtain actual performance target function value J (k) in the k moment；The target function revising forecast model network weight is:

g (k) = \frac{1}{2} {[J^{'} (k) - J (k)]}^{2} - - - (9)

Weighed value adjusting publicity is:

{ΔW}_{P}^{o u t} (k) = - η_{P} (k) {(\frac{\partial g (k)}{\partial W_{P}^{o u t} (k)})}^{T} - - - (10)

Wherein, η_PFor learning rate；

Step 2.3 optimization neural network uses ESN equally, and the input layer of optimization, output layer, dynamic reserve pool DR comprise respectively The individual neuron of K ', L ', N ', W_o ⁱⁿFor optimizing the ESN input layer connection weight value matrix to internal state reserve pool, W_oFor optimizing ESN Dynamically the weight matrix within reserve pool, is partially connected between its neuron, and degree of rarefication keeps 2%～5% equally, it is ensured that net Network has abundant dynamic characteristic, W_o ^backFor optimizing the ESN output layer connection weight value matrix to internal state reserve pool, W_o ⁱⁿ、 W_o、W_o ^backDimension is respectively N ' × K ', N ' × N ', N ' × L ', and initial value is and randomly generates, and once generation, the most no longer changes, The power spectral radius optimizing ESN should be less than 1 equally, and to ensure the stability of network, f is taken as Sigmoid function；

Its input is

H (k+1)=[R^T(k-1),S(k)]^T (11)

R^T(k-1) being k-1 moment dissolved oxygen concentration and the setting value of nitrate, S (k) is the ambient condition in k moment, represents For:

S (k)=[S_O(k),S_NO(k),S_NH(k),S_ND(k),X_ND(k),Q₀(k)] (12)

S_OK () is intake in the k moment dissolved oxygen concentration, S_NOK () is intake in the k moment nitrate, S_NHK () is ammonia of intaking in the k moment Concentration, S_NDK () is intake in the k moment soluble organic nitrogen, X_NDK () is intake in the k moment insolubility organic nitrogen, Q₀K () is the k moment Flow of inlet water；

The dynamic reserve pool optimizing ESN is output as:

x_{o} (k + 1) = f (W_{o}^{i n} h (k + 1) + W_{o} x_{o} (k) + W_{o}^{b a c k} R (k)) - - - (13)

ESN is output as setting value:

R (k + 1) = f^{o u t} (W_{o}^{o u t} x_{o} (k + 1)) - - - (14)

W_o ^outFor optimizing the ESN internal state reserve pool connection weight value matrix to output layer, dimension is L ' × N ', at learning process Middle needs are adjusted, f^outIt is taken as linear function；

R (k+1) is input in performance indications forecast model by step 2.4, obtains predicting that dynamic reserve pool new for ESN is output as:

x_{P} (k + 1) = f (W_{P}^{i n} R (k + 1) + W_{P} x_{P} (k) + W_{P}^{b a c k} J^{'} (k)) - - - (15)

Then performance indications predictive value during k+1 is:

J^{'} (k + 1) = f^{o u t} (W_{P}^{o u t} x_{P} (k + 1)) - - - (16)

It is as follows that step 2.5 carries out weighed value adjusting formula to optimization ESN:

{ΔW}_{o}^{o u t} (k + 1) = - η (k) {(\frac{\partial J^{'} (k + 1)}{\partial W_{o}^{o u t} (k + 1)})}^{T} - - - (17)

Wherein η (k) is learning rate；

After step 2.6 will optimize, newly generated setting value is input in PID controller as the value in k+1 moment, obtains at sewage In the reason process k+1 moment actual output J (k+1), performance indications forecast model weights are readjusted, the target function of weighed value adjusting For:

g (k + 1) = \frac{1}{2} {[J^{'} (k + 1) - J (k + 1)]}^{2} - - - (18)

3) k in step 2.1 to step 2.6 is added n, become k+n, repeat step 2.1 to step 2.6, n=1,2,3,4 ..., enter Row circulation, until no longer having into water data.