A data-driven model predictive control method for oxygen content of flue gas in municipal solid waste incineration process

Abstract

The present application provides a kind of data-driven urban solid waste incineration process flue gas oxygen content model predictive control method.For the inherent nonlinearity and uncertainty of solid waste incineration process, the characteristics of which cannot establish an accurate mathematical model, first, a prediction model based on self-organizing long short-term memory neural network is constructed, which can dynamically and adaptively adjust the number of hidden layer neurons, obtain a simplified model structure and higher flue gas oxygen content prediction accuracy.Secondly, an error feedback correction model based on long short-term memory neural network is established to correct the flue gas oxygen content prediction results, further improving the prediction accuracy of the model.Finally, a model predictive controller based on double long short-term memory neural network is designed, and the gradient descent method is used to roll optimization of the objective function of model predictive control, and the multi-step control law is efficiently solved.This method can better track the set value of flue gas oxygen content, and realize the accurate control of flue gas oxygen content in urban solid waste incineration process.

Description

TECHNICAL FIELD

[0001] The present application establishes a data-driven municipal solid waste incineration process flue gas oxygen content model predictive control method by using data-driven technology and model predictive control technology. Predictive control is one of the main development trends of industrial process intelligent control research and an important branch of advanced manufacturing technology field. The present application belongs to both the field of municipal solid waste incineration and the technical field of industrial process intelligent control research. BACKGROUND

[0002] With the rapid economic development, the improvement of residents' living standards and the acceleration of urbanization, about 2 billion tons of municipal solid waste are generated globally every year, and it is estimated that it will reach 25.9 billion tons by 2030. The problem of municipal solid waste treatment is increasingly prominent. At present, for municipal solid waste that cannot be directly recycled, it is generally treated by sanitary landfill, composting and incineration. Compared with other treatment methods, municipal solid waste incineration technology has the advantages of harmlessness, volume reduction and resource utilization, and represents the development trend of municipal solid waste treatment technology. The municipal solid waste incineration process is a large and complex system, including multiple subsystems and variable operating conditions. This process has complex operating mechanism, strong nonlinearity and random disturbance. The flue gas oxygen content at the horizontal flue outlet is a process parameter for measuring the incineration effect, and is closely related to the incineration efficiency and pollutant emission. If the flue gas oxygen content is too small, it indicates that the incomplete combustion heat loss increases, and the combustion efficiency will decrease, and a large amount of toxic and harmful gases such as dioxin, carbon monoxide and sulfur dioxide will be produced; if the flue gas oxygen content is too large, it indicates that the air amount is too large, and the excess air will take away a large amount of heat and dust, and also increase the generation of fuel-type nitrogen oxide NOx pollutants. Therefore, efficient and stable flue gas oxygen content control technology has always been the research focus of municipal solid waste incineration process control.

[0003] Model predictive control (MPC) is an advanced control algorithm for handling constrained optimization problems, which has certain self-adaptability and robustness, and is favored by many researchers and engineering technicians. However, one of the key factors restricting the practical application of model predictive control technology is the accurate modeling of nonlinear system prediction model. Data-driven modeling method is based on process data collection, which does not need to deeply understand the process mechanism and has strong algorithm universality. Among many data-driven models, long short-term memory network is a special recurrent neural network that can learn the long-term dependence between input sequence data through the gating mechanism on the internal storage unit, and has good prediction performance for time series data. Determining the appropriate network structure is one of the key steps in constructing long short-term memory neural network model, and too small or too large structure will cause underfitting or overfitting problems of neural network. Therefore, how to effectively construct a long short-term memory neural network prediction model with compact structure and excellent prediction performance according to the task to be processed is still one of the key research problems of data-driven model predictive control.

[0004] In actual incineration process, due to changes such as equipment aging, coking dirt and external disturbance, the process model will also change over time, and the data-driven model may not be able to correctly predict the future dynamics of the system after the disturbance occurs, ultimately affecting the performance of model predictive control. The traditional method is to compensate the model prediction by using the measured error, but this error feedback correction method has obvious hysteresis and is difficult to adapt to the dynamic changes of complex nonlinear systems. In order to improve the robustness of the model predictive controller in the dynamic environment of the municipal solid waste incineration process, the error correction model can be constructed to correct the dynamic model prediction error of the system, and further improve the effect of model predictive control. SUMMARY

[0005] In view of the shortcomings of the existing method, the present application provides a data-driven municipal solid waste incineration process flue gas oxygen content model predictive control method. A data-driven municipal solid waste incineration process flue gas oxygen content model predictive control framework is proposed, a flue gas oxygen content prediction method based on self-organizing long short-term memory neural network is designed, an error feedback correction model based on long short-term memory neural network is established, and a gradient descent method is used to solve the control law. The control method has good flue gas oxygen content variable setpoint tracking performance.

[0006] (1) Design the objective function for flue gas oxygen content model predictive control in municipal solid waste incineration process:

[0007]

[0008] Where, r(t) = [r(t+1), r(t+2),..., r(t+H p ) ] is the flue gas oxygen content setpoint vector at time t, and T is the flue gas oxygen content variable setpoint vector at time t, and is the flue gas oxygen content prediction value vector at time t, Δu(t) = [Δu(t), Δu(t+1),..., Δu(t+H u -1)] T is the flue gas oxygen content control amount adjustment vector at time t. H p is the prediction time domain, H p ∈ [1, 10]; H u is the control time domain, H u ∈ [1, 5]; wherein H u ≤ H p . T is the transpose of the formula, ρ1 and ρ2 are control weight factors, ρ1 ∈ [0, 1], ρ2 ∈ [0, 1]. The advantage of the data-driven model predictive controller is the accurate prediction model, no lag error feedback correction and efficient control law solving.

[0009] (2) Design a self-organizing long short-term memory neural network to establish a prediction model for the flue gas oxygen content predictive controller. The hidden layer neuron structure of the model can grow or be pruned according to the changes in the training data, achieving model structure simplification while improving model generalization performance.

[0010] The input of the self-organizing long short-term memory neural network is x(t) = [y(t-1), y(t-2), u1(t-1), u1(t-2), u2(t-1), u2(t-2), u3(t-1), u3(t-2)], y(t-1) is the actual value of the flue gas oxygen content at t-1 time in the solid waste incineration process, y(t-2) is the actual value of the flue gas oxygen content at t-2 time in the solid waste incineration process, u1(t-1) is the primary air flow rate in the dry section of the grate at t-1 time in the solid waste incineration process, u1(t-2) is the primary air flow rate in the dry section of the grate at t-2 time in the solid waste incineration process, u2(t-1) is the primary air flow rate in the first combustion section of the grate at t-1 time in the solid waste incineration process, u2(t-2) is the primary air flow rate in the first combustion section of the grate at t-2 time in the solid waste incineration process, u3(t-1) is the secondary air flow rate at t-1 time in the solid waste incineration process, and u3(t-2) is the secondary air flow rate at t-2 time in the solid waste incineration process. The output of the self-organizing long short-term memory neural network is the flue gas oxygen content prediction value.

[0011] The LSTM neural network realizes information update of unit state and hidden state through a gating structure, and the specific operation formula is as follows

[0012] f(t) = σ(w fx x(t) + w fh h(t-1) + b f ) (2)

[0013] i(t) = σ(w ix x(t) + w ihh(t - 1) + b i ) (3)

[0014] o(t) = σ(w ox x(t) + w oh h(t - 1) + b o ) (4)

[0015]

[0016]

[0017] h(t) = o(t) Θ τ(c(t)) (7)

[0018] where w fx , w ix , w ox and are input weights; w fh , w ih , w oh and are weights of recurrent structure; b f , b i , b o and are corresponding biases; σ and τ represent nonlinear sigmoid and tanh functions. Vector dot product operation is denoted as Θ.

[0019] After forward propagation, the prediction output of the LSTM network is computed as

[0020] y p (t) = w out h(t) (8)

[0021] where w out is the output weight. The loss function of the neural network training process is defined as

[0022]

[0023] where y d (t) and y p (t) represent the desired output and the predicted output. Here, the weights and biases are updated by the backpropagation through time algorithm

[0024]

[0025]

[0026]

[0027]

[0028] where * represents a certain specific item in f, i, o and η1 is the learning rate. To ensure the convergence of the neural network, the learning rate η1 is in the range of

[0029]

[0030] where,

[0031] The structure growth pruning process of the SOLSTM neural network is described in detail as follows.

[0032] On the one hand, the input variables are changed by a certain proportion and then input to the model again, the change of neuron output is observed, and the neuron response to input change is evaluated using the neuron activity evaluation index NIV. The specific calculation process is as follows:

[0033] ① Initialize the network model structure, input variable I j is increased and decreased by α proportion (α ∈ (0, 1)) in turn, to obtain two new input variables I j1 and I j2 .

[0034] I j1 = (1 + α)I j (15)

[0035] I j2 = (1 - α)I j (16)

[0036] ② The newly obtained input variables I j1 and I j2 are input through the network model respectively, to obtain two new neuron outputs A j1 and A j2 , the absolute value of the difference between the two is the response change of the neuron after the input variable is increased and decreased by α proportion, denoted as NIV j .

[0037] NIV j = |A j1 -A j2 | (17)

[0038] ③ Since the neuron influence value is a specific embodiment of the neuron activity, it is necessary to retain neurons with larger NIV values. Here, the cumulative NIV activity of the first m neurons is set as γ, defined as

[0039]

[0040] Where M is the initial number of hidden layer neurons. By setting the cumulative NIV activity threshold ξ, the most active top proportion (Ф∈(0, 1)) of neurons and the neurons with cumulative NIV activity rate γ higher than ξ are retained.

[0041] ④ Inhibiting the neurons that are insensitive to the changes of input variables, i.e. setting their neuron connection weight mask MASK to 0.

[0042] On the other hand, like most neural networks, the output layer of the LSTM network plays the role of summation. If the absolute value of the output weight is large, it means that the hidden layer neuron connected by the weight contributes more to the total output of the network. In order to improve the situation of over-pruning of neurons, these hidden layer neurons can be reactivated. Therefore, a significance indicator SI based on the output weight is proposed to evaluate the contribution of each hidden layer neuron.

[0043]

[0044] The larger the SI value, the more significant the contribution of the corresponding hidden layer neuron to the output. Therefore, according to the contribution significance of each LSTM neuron, the top proportion (β∈(0, 1)) of most significant neurons are found and activated, i.e. their connection weight mask MASK is set to 1.

[0045] (3) Designing a long short-term memory neural network to establish a prediction correction model to compensate for the prediction error.

[0046] The long short-term memory neural network is used to predict the system error online. The historical output y i (t) of the system, the control variable u i (t), and the historical deviation [e i (t), e i (t-1), …, e i (t-t l )] are used as the input of the neural network, and the error at the current time of the system is used as the output of the training sample, so that the output of the neural network is constantly approaching the actual error of the system, thereby establishing a prediction model of the system error.

[0047]

[0048] Where t l is the lag step. Since the long short-term memory neural network has strong time series data prediction ability, the trained network can be used as a prediction error correction model to predict the current time error using the historical input and output of the system and the historical error information, and then compensate for the influence of the prediction error of the flue gas oxygen content prediction model.

[0049]

[0050] The construction method and steps (2) of the long short-term memory neural network prediction correction model are similar.

[0051] (4) The gradient descent method is used to roll optimization model predictive control target function, efficient solution multi-step control law:

[0052]

[0053] Wherein, η2 is learning rate. Similarly, to ensure the convergence of the data-driven model predictive controller, the learning rate should also be limited to the value range according to formula (14). Take the first value of Δu(t) as the adjustment vector of the controller, that is, adjust the primary air flow of the dry section of the grate, the primary air flow of the grate combustion 1 section and the secondary air flow at t time:

[0054] u(t+1) = u(t) + Δu(t) (23)

[0055] Wherein, u(t+1) is the primary air flow of the dry section of the grate, the primary air flow of the grate combustion 1 section and the secondary air flow at t+1 time.

[0056] (5) The solved u(t) is used to control the oxygen content of flue gas, and the purpose of controlling the primary air flow of the dry section of the grate, the primary air flow of the grate combustion 1 section and the secondary air flow is achieved by adjusting the opening of the fan inlet baffle valve, and finally the stable and efficient control of the oxygen content of flue gas in municipal solid waste incineration process is realized.

[0057] Compared with the prior art, the present application has the following obvious advantages and beneficial effects:

[0058] 1. The present application proposes a flue gas oxygen content prediction model based on self-organizing long short-term memory neural network, which can dynamically and adaptively adjust the number of hidden layer neurons to obtain a simplified model structure and higher prediction accuracy.

[0059] 2. The present application proposes an error feedback correction model based on long short-term memory neural network, which compensates the prediction results, and solves the model mismatch problem caused by nonlinear disturbance or unmodeled dynamics.

[0060] 3. The present application proposes a data-driven model predictive control method, which realizes the stable tracking control of flue gas oxygen content. DETAILED DESCRIPTION

[0061] Figure 1 The data-driven model predictive control method of the present application;

[0062] Figure 2 The long short-term memory neural network unit structure diagram of the present application;

[0063] Figure 3 The self-organizing long short-term memory network neuron number change graph of the application;

[0064] Figure 4 The flue gas oxygen content prediction result graph of the application;

[0065] Figure 5 The flue gas oxygen content prediction error graph of the application;

[0066] Figure 6 The flue gas oxygen content control result graph of the application;

[0067] Figure 7 The flue gas oxygen content control error graph of the application. DETAILED DESCRIPTION

[0068] The application uses historical data stored by a distributed control system of a certain municipal solid waste incineration plant to verify and analyze the proposed modeling and control method. The application adopts the following technical solutions and implementation steps:

[0069] (1) Constructing a flue gas oxygen content prediction model based on a self-organizing long short-term memory neural network

[0070] The data used in the modeling stage comes from 1748 groups of measured data collected by the field distributed control system with a sampling interval of 2s. The first 1224 groups of data (70%) are selected as training samples, and the last 524 groups of data (30%) are selected as test samples to establish an offline prediction model based on a self-organizing long short-term memory neural network. Before control, the maximum lag n y = n u = 2, then x(t) = [y(t-1), y(t-2), u1(t-1), u1(t-2), u2(t-1), u2(t-2), u3(t-1), u3(t-2)].

[0071] The initial number of neurons of the self-organizing long short-term memory neural network is set to 30, the learning rate η1 = 0.05, the initial weights are randomly set, and the parameters Ф, α, ξ and β of the structure self-organizing process are set to 50%, 10%, 98% and 80%, respectively. The network is trained using the time backpropagation algorithm. The neuron number change is as shown in Figure 3 To verify the effectiveness of the network model, radial basis function neural network (RBF), fuzzy neural network (FNN) and LSTM (long short-term memory) neural network are selected for experimental comparison. The specific prediction curves of different methods are as shown in Figure 4 the prediction error curves are as shown inFigure 5 It can be seen that the proposed self-organizing long short-term memory neural network can better fit the real value curve of flue gas oxygen content and has better prediction accuracy.

[0072] In order to further compare the prediction effects of different models on flue gas oxygen content, root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE) are used to evaluate the prediction performance.

[0073]

[0074]

[0075]

[0076] where y d (t), y p (t) and Q are the real value, the predicted value and the number of test samples, respectively. The comparison of prediction performance of different methods is shown in Table 1. Compared with RBF and FNN, the values of RMSE, MAPE and MAE of LSTM and SOLSTM neural networks are lower, which reflects the advantage of LSTM network in solving this type of prediction problem. In addition, after the growth pruning process, the number of neurons of the final SOLSTM network is 10. Compared with the LSTM network, the structure is more compact. Therefore, the SOLSTM network is more suitable for simulating the dynamic characteristics of the change of flue gas oxygen content in the process of municipal solid waste incineration, and provides an accurate prediction model for the design of subsequent model predictive control scheme.

[0077] Table 1 Comparison of prediction performance of different methods

[0078]

[0079] (2) Error feedback correction model based on long short-term memory neural network

[0080] The historical output y i (t), the control amount u i (t) and the historical deviation e i (t) of the system are used as the input of the neural network, and the error at the current time of the system is used as the output of the training sample, so that the output of the neural network continuously approaches the actual error of the system, thereby establishing a prediction model of the system error.

[0081]

[0082] where t lis the hysteresis step. Since the LSTM neural network has strong time series data prediction ability, the trained network can be used as a prediction error correction model to predict the current time error using the system's historical input and output and historical error information, and then compensate for the influence of the prediction error of the flue gas oxygen content prediction model.

[0083]

[0084] (3) The DLSTM-MPC method is used to control the flue gas oxygen content.

[0085] Before the control strategy is implemented, some prediction control parameters need to be set in advance. The prediction horizon H p = 5, the control horizon H u = 1, the reference trajectory adjustment factor a r = 0.1, and the control weight factors p1 and p2 are 5 and 1. Figure 6 The set value tracking results of the flue gas oxygen content are shown in Fig. 6. It can be seen that the DLSTM-MPC can track the changing optimal set value. The tracking error of the flue gas oxygen content is shown in Fig. 7. Figure 7

[0086] In addition, in order to verify the effectiveness of the proposed controller, RBF, FNN and LSTM neural networks are selected as data-driven MPCs for the prediction model as comparative experiments. The absolute error integral (IAE) and the integral time absolute error (ITAE) are used to evaluate the control effect, and the specific formulas are as follows:

[0087]

[0088]

[0089] Table 2 Comparison of control performance of different strategies

[0090]

[0091] In order to accurately describe the performance of the proposed control method, Table 2 compares the control results of different methods. Compared with other neural network model predictive control methods, the proposed DLSTM-MPC control strategy has relatively low IAE and ITAE values. The results show that the proposed control method can better track the flue gas oxygen content set value and achieve precise control of the flue gas oxygen content.​

Claims

1. A data-driven model predictive control method for flue gas oxygen content of municipal solid waste incineration process, taking the primary air flow of the grate drying section, the primary air flow of the grate combustion 1 section, and the secondary air flow as the control variables, and the flue gas oxygen content as the controlled variable, characterized in that Comprising the following steps: (1) Design the objective function for the model predictive control of oxygen content in flue gas in the municipal solid waste incineration process: wherein r(t) = [r(t+1), r(t+2),..., r(t+H p )] T is the flue gas oxygen content set value vector at time t, is the flue gas oxygen content prediction value vector at time t, Δu(t) = [Δu(t), Δu(t+1),..., Δu(t+H u -1)] T is the flue gas oxygen content control amount adjustment vector at time t; H p is the prediction time domain, H p ∈ [1, 10]; H u is the control time domain, H u ∈ [1, 5]; wherein H u ≤ H p ; T is the transpose of the formula, ρ1 and ρ2 are control weight factors, ρ1 ∈ [0, 1], ρ2 ∈ [0, 1]; (2) Design a self-organizing long short-term memory neural network to establish a prediction model for the oxygen content in flue gas predictive controller; The input of the self-organizing long short-term memory neural network is x(t) = [y(t-1), y(t-2), u1(t-1), u1(t-2), u2(t-1), u2(t-2), u3(t-1), u3(t-2)], y(t-1) is the actual value of the oxygen content in flue gas in the municipal solid waste incineration process at t-1, y(t-2) is the actual value of the oxygen content in flue gas in the municipal solid waste incineration process at t-2, u1(t-1) is the primary air flow in the dry section of the grate in the municipal solid waste incineration process at t-1, u1(t-2) is the primary air flow in the dry section of the grate in the municipal solid waste incineration process at t-2, u2(t-1) is the primary air flow in the first combustion section of the grate in the municipal solid waste incineration process at t-1, u2(t-2) is the primary air flow in the first combustion section of the grate in the municipal solid waste incineration process at t-2, u3(t-1) is the secondary air flow in the municipal solid waste incineration process at t-1, and u3(t-2) is the secondary air flow in the municipal solid waste incineration process at t-2; The output of the self-organizing long short-term memory neural network is the predicted value of the oxygen content in flue gas; The LSTM neural network realizes the information update of the unit state and the hidden state through the gating structure, and the specific operation formula is as follows f(t) = σ(w fx x(t) + w fh h(t - 1) + b f ) (2) i(t) = σ(w ix x(t) + w ih h(t - 1) + b i ) (3) o(t) = σ(w ox x(t) + w oh h(t - 1) + b o ) (4) After forward propagation, the prediction output of the LSTM network is calculated as Where w fx w ix w ox and It is the input weight; w fh w ih w oh and It is the weight of the loop structure; b f b i b o and These are the corresponding biases; σ and τ represent the nonlinear sigmoid function and tanh function; the vector dot product is represented as ⊙; The structure growth and pruning process of the self-organizing long short-term memory neural network SOLSTM is described in detail as follows; y p (t) = w out h(t) (8) where w out is the output weight; the loss function of the neural network training process is defined as where y d (t) and y p (t) represent the desired output and the predicted output; here, the weights and biases are updated by a time backpropagation algorithm where * denotes a particular item in f, i, o and η1 is the learning rate; to ensure the convergence of the neural network, the learning rate η1 is in the range of wherein, On the one hand, the input variables are changed by increasing and decreasing according to a certain proportion and then input to the model again, the change of neuron output is observed, and the neuron response to input change is evaluated using the neuron activity evaluation index NIV; The specific calculation process is as follows: ③ Set the cumulative NIV activity of the first m neurons as γ, defined as ① Initialize the network model structure, and input variable I j According to the proportional increase and decrease of α, α∈(0, 1), two new input variables I j1 and I j2 are obtained. I j1 = (1 + a)I j (15) I j2 = (1 - a)I j (16) ②The newly obtained input variable I j1 and I j2 are respectively passed through the network model to obtain two new neuron outputs A j1 and A j2 , and the absolute value of the difference between the two is the response change of the neuron to the input variable after the proportion of α is increased or decreased, denoted as NIV j ; NIV j = | A j1 - A j2 | (17) Where M is the initial number of hidden layer neurons; By setting the cumulative NIV activity threshold ξ, the most active first Φ proportion of neurons and the neurons with cumulative NIV activity rate γ higher than ξ are retained; Φ ∈ (0, 1) ④ Suppress the remaining neurons that are not sensitive to the change of input variables, i.e. set their neuron connection weight mask MASK to 0; A significance index SI based on output weight is proposed to evaluate the contribution of each hidden layer neuron; The greater the SI value, the more significant the contribution of the corresponding hidden layer neuron to the output; Therefore, according to the contribution significance of each LSTM neuron, the most significant first β proportion (β ∈ (0, 1)) of neurons are found out and activated; That is, their connection weight mask MASK is set to 1; (3) Design a long short-term memory neural network to establish a prediction correction model to compensate for the prediction error; (4) Use gradient descent method to rollingly optimize the objective function of model predictive control to solve the multi-step control law: The system error is predicted on line using a long short-term memory neural network; the historical output y i (t) of the system i (t) and the historical deviation [e i (t), e i (t-1), e i (t-t l )] of the system are used as the training sample input of the neural network, and the error of the current time of the system is used as the training sample output, so that the output of the neural network continuously approaches the actual error of the system, thereby establishing a prediction model of the system error; where t l is the lag step; since the long short-term memory neural network has strong time series data prediction ability, the trained network can be used as a prediction error correction model to predict the current time error using the system's historical input and output and historical error information, and then compensate for the influence of the flue gas oxygen content prediction model prediction error. ​ Wherein, η2 is the learning rate; similarly, to ensure the convergence of the data-driven model predictive controller, the learning rate should also be limited in the range of formula (14); take the first value of Δu(t) as the adjustment vector of the controller, that is, adjust the primary air flow of the drying section of the grate, the primary air flow of the first combustion section of the grate and the secondary air flow at time t: u(t+1) = u(t) + Δu(t) (23) Wherein, u(t+1) is the primary air flow of the drying section of the grate, the primary air flow of the first combustion section of the grate and the secondary air flow at time t+1; (5) Control the oxygen content in the flue gas by using the solved u(t), adjust the opening degree of the fan inlet baffle valve to control the primary air flow of the drying section of the grate, the primary air flow of the first combustion section of the grate and the secondary air flow, and finally realize the control of the oxygen content in the flue gas of the municipal solid waste incineration process.

Images