Municipal sewage treatment supply chain product quality frequency domain topological feature prediction method

Abstract

The invention provides an urban sewage treatment supply chain product quality frequency domain topological feature prediction method. The method comprises the following steps: collecting operation data such as flow, oxidation-reduction potential, nitrate nitrogen concentration, dissolved oxygen concentration and effluent quality as input variables through a multi-source sensor, and establishing a prediction model based on a frequency domain topological organization neural network; constructing a frequency domain feature embedding layer and a frequency domain topological feature coding layer, performing frequency domain causal feature extraction on input variables, and capturing a topological dependence and propagation delay relationship between the variables; a frequency domain topological organization attention mechanism is designed, a frequency domain causal delay matrix is used for organizing and forming a frequency domain connection strength matrix, and the frequency domain connection strength matrix is used for adjusting the connection weight of a query matrix and a key matrix in the attention mechanism; the method can weaken the influence of multi-path propagation interference on a prediction result, and realizes high-precision prediction of product quality indexes such as total nitrogen and total phosphorus in the effluent of the urban sewage treatment supply chain.

Description

TECHNICAL FIELD

[0001] The present application designs a municipal sewage treatment supply chain product quality frequency domain topological feature prediction method, which realizes accurate prediction of key product quality indexes such as total nitrogen concentration of effluent and total phosphorus concentration of effluent in the municipal sewage treatment process. The product quality of the municipal sewage treatment supply chain is an important parameter reflecting the purification effect and effluent standard level of the treatment process, which can directly reflect the stability of process operation, pollutant removal efficiency and system energy consumption, and is an important basis for formulating operation control strategy and optimizing process control. The municipal sewage treatment plant is a typical complex process industrial system, which is widely used in the fields of urban water resource recycling and pollutant reduction. Therefore, predicting the product quality of the municipal sewage treatment supply chain is of great significance to improve the operation management level, ensure the process stability and optimize the energy consumption distribution, and belongs to the fields of sewage treatment supply chain management and intelligent prediction technology. BACKGROUND

[0002] The municipal sewage treatment supply chain is a typical process industrial supply chain, which is widely used in the fields of municipal sewage treatment, water resource recycling and advanced sewage treatment, and undertakes key functions such as material transportation, reaction control and effluent quality guarantee in the sewage treatment process. The change of product quality of the municipal sewage treatment supply chain directly reflects the purification effect and effluent standard level of the system, and is one of the important bases for evaluating process operation state and formulating optimization control strategy. If the change of product quality of the municipal sewage treatment supply chain is not predicted in time and accurately, it may lead to unreasonable adjustment of reaction zone operation parameters, intensified effluent water quality fluctuation, even cause system load imbalance and energy consumption increase, and affect the stability and standard discharge effect of the treatment system. Therefore, accurate prediction of the product quality of the municipal sewage treatment supply chain is an important means to ensure stable operation of the system and realize energy saving and emission reduction.

[0003] In the prior art, the prediction method for the product quality of the municipal sewage treatment supply chain mainly relies on static modeling, empirical formula or single variable time series analysis, which is difficult to accurately depict the complex dynamic coupling relationship between multiple variables and cannot effectively represent the topological correlation characteristics between different reaction zones. Under the conditions of multiple working condition switching, fluctuation of influent water quality and frequent external disturbance, the dependent structure between variables evolves with time, and the propagation path presents nonlinear and time delay characteristics, so that the traditional time domain prediction model is difficult to capture the multi-path propagation law in the frequency domain, the prediction result is easy to deviate due to noise interference, and the stability and generalization ability of the model are limited.

[0004] This invention designs a frequency domain topological feature prediction method for product quality in the urban wastewater treatment supply chain. This method uses sensors to collect multi-source data such as flow rate, water quality, and reaction zone characteristics during the operation of urban wastewater treatment, establishes a prediction model based on a frequency domain topological organization neural network, mines the topological dependencies between variables through frequency domain causal modeling, and reveals multipath propagation characteristics by using frequency phase expansion and delay matrix calculation. This enables accurate prediction of key product quality indicators such as total nitrogen and total phosphorus in effluent, significantly improving the prediction accuracy and operational stability of the urban wastewater treatment supply chain under complex operating conditions. Summary of the Invention

[0005] This invention provides a frequency-domain topological feature prediction method for product quality in the urban wastewater treatment supply chain. The method uses multi-source operational data collected by sensors during the urban wastewater treatment process, including influent flow rate, external return flow rate, internal return flow rate, aeration rate, redox potential in anaerobic and anoxic zones, dissolved oxygen concentration in aerobic zones, and effluent quality, as input variables to establish a product quality prediction model based on a frequency-domain topological organization neural network. The model captures the topological dependencies between variables through frequency-domain causal modeling and identifies multipath propagation delay characteristics using a frequency phase expansion mechanism. It also dynamically adjusts the connection weights between variables using a frequency-domain topological attention structure, thereby mitigating the impact of complex coupling relationships and propagation interference on prediction accuracy. Through online training and dynamic updating of the model parameters, a prediction model with time-varying dependency adaptation and topological interference suppression capabilities is formed, achieving high-precision prediction of future product quality in the urban wastewater treatment supply chain.

[0006] The present invention adopts the following technical solution and implementation steps:

[0007] 1. A method for predicting the frequency domain topological features of urban wastewater treatment supply chain products, characterized by: collecting urban wastewater treatment operation data, establishing a frequency domain topological feature prediction model for urban wastewater treatment supply chain products, training the frequency domain topological feature prediction model for urban wastewater treatment supply chain products, and predicting the product quality of urban wastewater treatment supply chain products, comprising the following steps:

[0008] (1) Collect urban wastewater treatment operation data:

[0009] With a municipal sewage treatment plant as the research object, the operation process of the municipal sewage treatment plant includes water inlet transportation, external reflux, internal reflux, anaerobic reaction, anoxic reaction, aerobic reaction, sludge discharge and effluent discharge links, and each link is sequentially connected to constitute an operation unit of a municipal sewage treatment supply chain; the treated effluent is taken as a product of the municipal sewage treatment supply chain, and the product quality is determined by the effluent quality, the effluent quality takes the total nitrogen concentration and the total phosphorus concentration of the effluent as main indexes, and is used for reflecting the purification effect and the effluent standard level of the municipal sewage treatment supply chain; multi-source operation data reflecting flow, water quality and reaction zone characteristics are collected, an electromagnetic flowmeter installed at an outlet pipeline of a water inlet pump house is used to collect an inlet flow rate x 1 (t) at t time, a unit is cubic meters / hour, an electromagnetic flowmeter installed at an outlet pipeline of an external reflux pump is used to collect an external reflux flow rate x 2 (t) at t time, a unit is cubic meters / hour, an online ammonia nitrogen analyzer installed at a branch pipeline of the water inlet pipeline is used to collect an inlet ammonia nitrogen concentration x 3 (t) at t time, a unit is milligrams / liter, a redox potential sensor installed inside an anaerobic tank is used to collect an anaerobic tank redox potential x 4 (t) at t time, a unit is millivolts, a redox potential sensor installed at an inlet of an anoxic zone is used to collect an anoxic zone inlet redox potential x 5 (t) at t time, a unit is millivolts, an online nitrate nitrogen analyzer installed at a reaction tank of the anoxic zone is used to collect an anoxic zone nitrate nitrogen concentration x 6 (t) at t time, a unit is milligrams / liter, a suspended solid concentration meter installed at the reaction tank of the anoxic zone is used to collect an anoxic zone mixed liquid suspended solid concentration x 7 (t) at t time, a unit is milligrams / liter, an electromagnetic flowmeter installed at an outlet pipeline of an internal reflux pump is used to collect an internal reflux flow rate x 8 (t) at t time, a unit is cubic meters / hour, a dissolved oxygen sensor installed at a front section of an aerobic zone is used to collect an aerobic zone front section dissolved oxygen concentration x 9 (t) at t time, a unit is milligrams / liter, a dissolved oxygen sensor installed at a rear section of the aerobic zone is used to collect an aerobic zone rear section dissolved oxygen concentration x 10 (t) at t time, a unit is milligrams / liter, a thermal type mass flowmeter installed at a first branch pipeline of an aeration pipe is used to collect a first aeration amount x 11 (t) at t time, a unit is cubic meters / hour, a thermal type mass flowmeter installed at a second branch pipeline of the aeration pipe is used to collect a second aeration amount x 12 (t) at t time, a unit is cubic meters / hour, a thermal type mass flowmeter installed at a third branch pipeline of the aeration pipe is used to collect a third aeration amount x 13 (t) at t time, a unit is cubic meters / hour, an electromagnetic flowmeter installed at a residual sludge discharge pipeline is used to collect a sludge discharge amount x 14(t), unit: cubic meters / hour, the total nitrogen concentration x of the effluent at time t is collected by an online total nitrogen analyzer installed in the effluent pipeline 15 (t), unit: milligrams / liter, the total phosphorus concentration x of the effluent at time t is collected by an online total phosphorus analyzer installed in the effluent pipeline 16 (t), unit: milligrams / liter; x 1 (t), x 2 (t), x 3 (t), x 4 (t), x 5 (t), x 6 (t), x 7 (t), x 8 (t), x 9 (t), x 10 (t), x 11 (t), x 12 (t), x 13 (t), x 14 (t), x 15 (t), x 16 (t) as the input variable of the product quality prediction model of the municipal sewage treatment supply chain at time t, the effluent total nitrogen concentration and the effluent total phosphorus concentration at the next time are the output variables of the model; t represents the current time step, N is the number of sample sequences, t = 1, 2, …, N;

[0010] (2) Establish a frequency domain topological feature prediction model of the product quality of the municipal sewage treatment supply chain:

[0011] A frequency domain topological organization neural network-based product quality frequency domain topological feature prediction model of the municipal sewage treatment supply chain is constructed, which is composed of an input layer, a frequency domain feature embedding layer, a frequency domain topological feature coding layer, and an output layer.

[0012] Input layer: The collected data at time t and its previous L-1 historical time steps are arranged in time sequence to form an input variable vector, X n (t) represents the nth input variable vector, X n (t) = [x n (t-L+1), …, x n (t)], x n (t-L+1) represents the value of the nth input variable at time step t-L+1, L is the length of the input variable time window, n represents the index of the input variable, n = 1, 2, …, 16; the output matrix of the input layer is X(t) = [X 1 (t) T ,…, X n (t) T ,…, X 6 (t)T ] T , T represents the transpose of a matrix;

[0013] The frequency domain feature embedding layer: the input layer output matrix is mapped to the frequency domain by discrete Fourier transform, and X(t) is decomposed into K=L / 2 frequency spectrum matrices; inverse Fourier transform is performed on each frequency spectrum matrix to obtain the corresponding time domain feature matrix, and the time domain feature matrix is stacked in the frequency dimension to form the frequency domain feature token tensor T1(t), wherein T(t)=[T 1,1 (t),...,T 1,k (t),...,T 1,K (t)], T 1,k (t) represents the frequency domain feature token matrix output by the frequency domain feature embedding layer at the kth frequency, k=1, 2, …, K;

[0014] The frequency domain topology feature encoding layer: composed of 3 encoding modules, s=1, 2, 3, s represents the number of the s-th encoding module; the influence weight of the historical time step on the current time step is calculated by the least square estimation method, and a multivariate autoregressive model is constructed to represent the time sequence dependence relationship between nodes; the autoregressive model is mapped to the frequency domain space based on the power spectral density function, and the frequency domain transfer function matrix is calculated at K discrete frequency points; the frequency domain causal topology matrix C k (t) represents the frequency domain causal topology matrix at frequency k at time t, and the phase delay difference between adjacent frequency points is calculated based on the frequency phase expansion of the frequency domain transfer function matrix, and the frequency domain causal delay matrix τ k (t) represents the frequency domain causal delay matrix at frequency k at time t; the single-step frequency domain topology causal matrix and the single-step frequency domain topology delay matrix are calculated:

[0015]

[0016] wherein, represents the single-step frequency domain topology causal matrix at frequency k at time t, represents the single-step frequency domain topology delay matrix at frequency k at time t, and represents the Hadamard product operator, 1 16×16 represents a full 1 matrix with a dimension of 16x16, I 16 represents a unit column vector with a dimension of 16, and represents a binary function for screening effective topology connections; the frequency domain causal matrix and the frequency domain causal delay matrix under multi-step topology propagation are calculated by recursion:

[0017]

[0018] wherein, denotes the i-th frequency-domain topological causal matrix of the r-th step at the k-th frequency point at the t-th time instant, denotes the i-th frequency-domain causal delay matrix of the r-th step at the k-th frequency point at the t-th time instant, denotes the i-th frequency-domain topological causal matrix of the r-1-th step at the k-th frequency point at the t-th time instant, q denotes the i-th frequency-domain topological causal matrix of the r-1-th step at the k-th frequency point at the t-th time instant, denotes the i-th frequency-domain causal delay matrix of the r-1-th step at the k-th frequency point at the t-th time instant, q denotes the i-th frequency-domain topological causal matrix of the r-1-th step at the k-th frequency point at the t-th time instant, denotes the i-th row vector of the extraction matrix, i m denotes the i-th row vector of the extraction matrix, i q denotes the quotient obtained by dividing the index i by 16, i m denotes the remainder obtained by dividing the index i by 16, r denotes the number of steps of topological propagation, R denotes the assumed propagation step length, r = 1, 2,..., R, i denotes the node index number, i = 1, 2,..., 16 r ; at each frequency point k, the frequency-domain delay perception value is calculated by weighted superposition of the delay matrix and the topological causal strength matrix of each step of topological propagation:

[0019]

[0020] wherein v k,l (t) denotes the frequency-domain delay perception value of the l-th time step at the k-th frequency point at the t-th time instant, [·] mn denotes the element of the m-th row and the n-th column of the extraction matrix, exp(·) denotes the exponential function operation, l is the time step index, denotes the time sequence position corresponding to different delay steps, m denotes the node row index, n denotes the node column index, l = -L + 1,..., L - 1, m = 1, 2,..., 16, n = 1, 2,..., 16; according to the frequency-domain delay perception result, the frequency-domain connection strength matrix is constructed:

[0021] [M k (t)] gh = v k,g-h (t), -L + 1 ≤ g - h ≤ L - 1 (6)

[0022] wherein M k (t) denotes the frequency-domain connection strength matrix at the k-th frequency point at the t-th time instant, when the time step difference g - h is within the effective range [-L + 1, L - 1], when the time step difference connection strength matrix element is derived from the delay perception value v k,g-h(t) = 0, otherwise, g denotes the connection strength matrix row index, h denotes the connection strength matrix column index, g = 1, 2, …, L, n = 1, 2, …, L; constructing a frequency domain topological organization attention, adjusting the connection strength between the query matrix and the key matrix through the frequency domain connection strength matrix, dynamically correcting the attention weight distribution, and making the attention allocation conform to the multi-path propagation relationship between the states of nodes; calculating the query matrix, the key matrix and the value matrix of the s th encoding module: Q s,k (t) = T s,k (t)W Q,s (t), K s (t) = T s,k (t)W K,s (t), V s (t) = T s,k (t)W V,s (t), Q s,k (t) denotes the query matrix of the s th encoding module under frequency k at time t, T s,k (t) denotes the input frequency domain feature token matrix of the s th encoding module under frequency k at time t, W Q,s (t) denotes the query weight matrix of the s th encoding module at time t, K s,k (t) denotes the key matrix of the s th encoding module under frequency k at time t, W K,s (t) denotes the key weight matrix of the s th encoding module, V s,k (t) denotes the value matrix of the s th encoding module under frequency k at time t, W V,s (t) denotes the value weight matrix of the s th encoding module at time t; calculating the frequency domain topological organization attention output matrix:

[0023] O s,k (t) = softmax(Q s,k (t)K s,k (t) T ⊙M k (t))V s,k (t) (7)

[0024] wherein, O s,k (t) denotes the frequency domain topological organization attention output matrix of the s th encoding module under frequency k at time t, softmax(·) denotes normalizing the matrix in the parentheses in the row dimension, so that the sum of each row element is 1; performing residual connection operation and layer normalization operation on the attention output matrix O s,k (t) of the s th encoding module under frequency k, to obtain the residual update output matrix R s,k (t) of the s th encoding module under frequency k; performing residual connection operation and layer normalization operation on the residual update output matrix R s,k(t) inputting the feedforward neural network to obtain the feedforward network output matrix of the s-th encoding module at frequency k The feedforward neural network structure comprises a first linear mapping layer with an input dimension of D and an output dimension of D, and a second linear mapping layer with an input dimension of D and an output dimension of D, and a nonlinear activation layer connected between the two, and the nonlinear activation function is a linear rectifier function; W 2,s (t) represents the first linear mapping weight matrix of the s-th encoding module, b 2,s (t) represents the first linear mapping bias vector of the s-th encoding module, W 3,s (t) represents the second linear mapping weight matrix of the s-th encoding module, b 3,s (t) represents the second linear mapping bias vector of the s-th encoding module; the feedforward network output matrix of the s-th encoding module at frequency k performing residual connection operation and layer normalization operation to obtain the frequency domain feature token matrix T output at frequency k of the s-th encoding module at time t s+1,k (t), represents the p-th row vector of the frequency domain feature token output matrix at frequency k of the s-th encoding module, p represents the index of the frequency domain feature token dimension, p = 1, 2,..., L;

[0025] Output layer: calculate the frequency domain feature prediction model output:

[0026]

[0027] wherein, represents the next time city sewage treatment supply chain product quality vector predicted by the model at time t, y1(t) represents the total nitrogen concentration of effluent predicted by the model at time t, y2(t) represents the total phosphorus concentration of effluent predicted by the model at time t, represents the row vector of the frequency domain feature token tensor unfolded by the output of the third encoding module, W4(t) represents the weight matrix of the output layer linear regression at time t, b4(t) represents the bias vector of the linear regression at time t;

[0028] (3) Training the city sewage treatment supply chain product quality frequency domain topological feature prediction model:

[0029] ① Define the loss function of the model as:

[0030]

[0031] wherein, J(t) represents the loss of the model at time t, y(t) represents the model-predicted next-time urban sewage treatment supply chain product quality vector at time t-1, y(t-1) represents the actual next-time urban sewage treatment supply chain product quality vector at time t-1, y(t-1) = [x 1 (t), x 2 (t)], that is, the actual next-time supply chain product quality at time t-1 is obtained from the actual data of the total nitrogen concentration and total phosphorus of the effluent collected at time t, and is used as a supervision signal at time t to participate in model training;

[0032] ② Set the current training time as t, initialize the training round number τ = 1, and set the training iteration number as 20. Initialize the weight parameters and bias parameters of the model, the weight parameters are randomly taken in the interval [-0.2, 0.2], and the bias parameter is taken as 0;

[0033] ③ Calculate the prediction output of the model of the τth training at time t-1 by using formulas (1)-(7) Calculate the loss J of the model of the τth training at time t by using formula (8) τ (t), update the weight matrix and the bias matrix by using the gradient descent method, and the calculation formula is:

[0034]

[0035] Among them, W τ (t) represents the model weight matrix set of the τth training at time t, including: the encoding module query weight matrix, the encoding module key weight matrix, the encoding module value weight matrix, the encoding module first linear mapping weight matrix, the encoding module second linear mapping weight matrix and the output layer linear regression weight matrix, J (t) represents the loss function J τ (t) of the τth training at time t, and the gradient of the model weight matrix set W τ (t) is W τ+1 (t) represents the model weight matrix set of the τ+1th training at time t, and b τ (t) represents the model bias vector set of the τth training at time t, including: the encoding module first linear mapping bias vector, the encoding module second linear mapping bias vector and the output layer linear regression bias vector, J (t) represents the loss function J τ (t) of the τth training at time t, and the gradient of the model bias vector set b τ (t) is b τ+1 (t) represents the model bias vector set of the τ+1th training at time t, τ represents the index of the training round number, τ = 1, 2, …, 20;

[0036] ④ If the number of training rounds τ < 20, increase τ by 1 and go to step ③ to continue training; if the number of training rounds τ ≥ 20, terminate the training and update of model parameters and go to step (4) to perform prediction.

[0037] (4) Predicting the quality of products in the urban wastewater treatment supply chain:

[0038] Using a trained urban wastewater treatment supply chain product quality frequency domain topological feature prediction model, the influent flow rate, external return flow rate, influent ammonia nitrogen concentration, anaerobic tank oxidation-reduction potential, anoxic zone inlet oxidation-reduction potential, anoxic zone nitrate nitrogen concentration, anoxic zone mixed liquor suspended solids concentration, internal return flow rate, dissolved oxygen concentration in the first section of the aerobic zone, dissolved oxygen concentration in the second section of the aerobic zone, first aeration rate, second aeration rate, third aeration rate, sludge discharge rate, effluent total nitrogen concentration, and effluent total phosphorus concentration collected at time t are used as model inputs. The model output vector at time t is obtained according to formulas (1)-(8). Model output vector As the predicted quality of urban wastewater treatment supply chain products at time t, the quality of urban wastewater treatment supply chain products includes: total nitrogen concentration in effluent (mg / L) and total phosphorus concentration in effluent (mg / L).

[0039] The inventiveness of this invention is mainly reflected in:

[0040] 1. This invention addresses the problem of complex topological dependencies and propagation path delay differences among multiple source variables during urban wastewater treatment operations, which leads to a decrease in the accuracy of supply chain product quality prediction. It proposes a frequency domain topological feature prediction method for urban wastewater treatment supply chain products. This method constructs a frequency domain feature embedding layer and a frequency domain topological feature encoding layer, performs frequency domain mapping and causal delay feature extraction on the input data, captures the time-varying topological dependencies and multi-path propagation delay patterns among variables, and achieves high-precision prediction of urban wastewater treatment supply chain product quality.

[0041] 2. This invention designs a frequency-domain topology-organized attention mechanism. Using a frequency-domain causal delay matrix as the core, it organizes a frequency-domain connection strength matrix to adjust the connection weights between the query matrix and the key matrix in the attention mechanism. This mechanism introduces topological propagation direction and delay features during attention allocation, enabling dynamic modeling of multi-path propagation dependencies. This effectively suppresses propagation interference between variables, enhances the model's ability to identify and maintain key topological features, and improves the accuracy and robustness of product quality prediction in the urban wastewater treatment supply chain. Attached Figure Description

[0042] Figure 1 This is a graph showing the prediction results of the total nitrogen concentration in the effluent from this invention;

[0043] Figure 2is a result figure of the total phosphorus concentration of the effluent predicted by the present application;

[0044] Figure 3 is an error figure of the total nitrogen concentration of the effluent predicted by the present application;

[0045] Figure 4 is an error figure of the total phosphorus concentration of the effluent predicted by the present application; DETAILED DESCRIPTION

[0046] The present application obtains a municipal sewage treatment supply chain product quality frequency domain topological feature prediction method, selects flow, oxidation-reduction potential, nitrate nitrogen concentration, dissolved oxygen concentration and effluent water quality and the like operation data as model input variables, and the total nitrogen concentration and the total phosphorus concentration of the effluent at the next moment as model outputs.

[0047] The experimental data is obtained by online sensor collection: a KROHNE OPTIFLUX 4300 type electromagnetic flowmeter is installed on the outlet pipeline of the influent pump house for collecting influent flow; an Endress+Hauser Proline Promag 10D type electromagnetic flowmeter is installed on the outlet pipeline of the external reflux pump for collecting external reflux flow; an Endress+Hauser Proline Promag W 400 type electromagnetic flowmeter is installed on the outlet pipeline of the internal reflux pump for collecting internal reflux flow; HACH ORP electrodes are installed on the anaerobic tank and the anoxic tank for collecting oxidation-reduction potential of each zone; a HACH Nitratax sc type online nitrate nitrogen analyzer is installed on the anoxic zone reaction tank for collecting nitrate nitrogen concentration; the mixed liquid suspended solid concentration is collected by a HACH SOLITAX sc type suspended solid concentration meter; a HACH LDO type dissolved oxygen sensor is installed on the aerobic zone for collecting dissolved oxygen concentration; a HACH TN sc type total nitrogen analyzer and a HACH TP sc type total phosphorus analyzer are installed on the effluent pipeline for collecting total nitrogen and total phosphorus concentrations of the effluent. The data sampling period is 15 minutes, and 400 groups of operation data samples are obtained by continuous collection for 50 hours; the data is divided into a training set and a test set in chronological order, wherein 200 groups of data in the first 25 hours are used as training samples, and 200 groups of data in the last 25 hours are used as test samples.

[0048] The present application adopts the following technical solutions and implementation steps:

[0049] A municipal sewage treatment supply chain product quality frequency domain topological feature prediction method, characterized by: collecting municipal sewage treatment operation data, establishing a municipal sewage treatment supply chain product quality frequency domain topological feature prediction model, training the municipal sewage treatment supply chain product quality frequency domain topological feature prediction model, and predicting the municipal sewage treatment supply chain product quality, comprising the following steps:

[0050] 1. Collecting urban sewage treatment operation data:

[0051] With urban sewage treatment plants as the research object, the operation process of the urban sewage treatment plant includes water inlet transportation, external reflux, internal reflux, anaerobic reaction, anoxic reaction, aerobic reaction, sludge discharge and effluent discharge links, which are sequentially connected to constitute the operation unit of the urban sewage treatment supply chain; the treated effluent is taken as the product of the urban sewage treatment supply chain, and the product quality is determined by the effluent quality, which takes the total nitrogen concentration and total phosphorus concentration of the effluent as the main indicators, and is used to reflect the purification effect and effluent standard level of the urban sewage treatment supply chain; multi-source operation data reflecting flow, water quality and reaction zone characteristics are collected, the electromagnetic flowmeter installed at the outlet pipeline of the water inlet pump house is used to collect the water inlet flow x 1 (t) at time t, with the unit of cubic meters / hour, the electromagnetic flowmeter installed at the outlet pipeline of the external reflux pump is used to collect the external reflux flow x 2 (t) at time t, with the unit of cubic meters / hour, the online ammonia nitrogen analyzer installed at the branch pipe of the water inlet pipeline is used to collect the water inlet ammonia nitrogen concentration x 3 (t) at time t, with the unit of milligrams / liter, the oxidation-reduction potential sensor installed inside the anaerobic tank is used to collect the oxidation-reduction potential of the anaerobic tank x 4 (t) at time t, with the unit of millivolts, the oxidation-reduction potential sensor installed at the inlet of the anoxic zone is used to collect the oxidation-reduction potential of the anoxic zone inlet x 5 (t) at time t, with the unit of millivolts, the online nitrate nitrogen analyzer installed at the anoxic zone reaction tank is used to collect the anoxic zone nitrate nitrogen concentration x 6 (t) at time t, with the unit of milligrams / liter, the suspended solid concentration meter installed at the anoxic zone reaction tank is used to collect the anoxic zone mixed liquid suspended solid concentration x 7 (t) at time t, with the unit of milligrams / liter, the electromagnetic flowmeter installed at the outlet pipeline of the internal reflux pump is used to collect the internal reflux flow x 8 (t) at time t, with the unit of cubic meters / hour, the dissolved oxygen sensor installed at the front section of the aerobic zone is used to collect the dissolved oxygen concentration x 9 (t) at the front section of the aerobic zone at time t, with the unit of milligrams / liter, the dissolved oxygen sensor installed at the rear section of the aerobic zone is used to collect the dissolved oxygen concentration x 10 (t) at time t, with the unit of milligrams / liter, the thermal mass flowmeter installed at the first branch pipe of the aeration pipe is used to collect the first aeration amount x 11 (t) at time t, with the unit of cubic meters / hour, the thermal mass flowmeter installed at the second branch pipe of the aeration pipe is used to collect the second aeration amount x 12 (t) at time t, with the unit of cubic meters / hour, the thermal mass flowmeter installed at the third branch pipe of the aeration pipe is used to collect the third aeration amount x 13(t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 14 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 15 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 16 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 1 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 2 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 3 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 4 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 5 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 6 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 7 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 8 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 9 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 10 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 11 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 12 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 13 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 14 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 15 (t), unit: cubic meters / hour, the sludge discharge amount x at time t is collected by using an electromagnetic flowmeter installed in the residual sludge discharge pipeline 16 (t) as the input variable of the municipal sewage treatment supply chain product quality prediction model at time t, the effluent total nitrogen concentration and the effluent total phosphorus concentration at the next time as the output variables of the model; t represents the current time step, N is the number of sample sequences, t = 1, 2, …, N;

[0052] 2. Establishing a municipal sewage treatment supply chain product quality frequency domain topology feature prediction model:

[0053] A municipal sewage treatment supply chain product quality frequency domain topology feature prediction model based on a frequency domain topology organized neural network is constructed, and the frequency domain topology organized neural network is composed of an input layer, a frequency domain feature embedding layer, a frequency domain topology feature coding layer, and an output layer.

[0054] Input layer: the collected data at time t and its previous L-1 historical time steps are arranged in time sequence to form an input variable vector, X n (t) represents the nth input variable vector, X n (t) = [x n (t-L+1), …, x n (t)], x n (t-L+1) represents the value of the nth input variable at time step t-L+1, L is the length of the input variable time window, n represents the index of the input variable, n = 1, 2, …, 16; the output matrix of the input layer is X 1 (t) T ,…, Xn (t) T ,…,X 6 (t) T ] T , T represents the transpose of a matrix;

[0055] Frequency domain feature embedding layer: the input layer output matrix is mapped to the frequency domain by discrete Fourier transform, X(t) is decomposed into K = L / 2 frequency spectrum matrices; inverse Fourier transform is performed on each frequency spectrum matrix to obtain the corresponding time domain feature matrix, and the time domain feature matrix is stacked in the frequency dimension to form a frequency domain feature token tensor T1(t), wherein T(t) = [T 1,1 (t),...,T 1,k (t),...,T 1,K (t)], T 1,k (t) represents the frequency domain feature token matrix output by the frequency domain feature embedding layer at the kth frequency, k = 1, 2, …, K;

[0056] Frequency domain topology feature encoding layer: composed of 3 encoding modules, s = 1, 2, 3, s represents the number of the s-th encoding module; the influence weight of the historical time step on the current time step is calculated by the least square estimation method, a multivariate autoregressive model is constructed to represent the time sequence dependence between nodes; the autoregressive model is mapped to the frequency domain space based on the power spectral density function, and the frequency domain transfer function matrix is calculated at K discrete frequency points; the frequency domain causal topology matrix C k (t) represents the frequency domain causal topology matrix at frequency k at time t, the phase delay difference between adjacent frequency points is calculated based on the frequency phase expansion of the frequency domain transfer function matrix, and the frequency domain causal delay matrix τ k (t) represents the frequency domain causal delay matrix at frequency k at time t; the single-step frequency domain topology causal matrix and the single-step frequency domain topology delay matrix are calculated:

[0057]

[0058] wherein, represents the single-step frequency domain topology causal matrix at frequency k at time t, represents the single-step frequency domain topology delay matrix at frequency k at time t, and represents the Hadamard product operator, 1 16×16 represents a 16x16 all-1 matrix, I 16 represents a 16-dimensional unit column vector, and represents a binary function for filtering effective topology connections; the frequency domain causal matrix and the frequency domain causal delay matrix under multi-step topology propagation are calculated by recursion:

[0059]

[0060] wherein, denotes the ith frequency domain topological causal matrix of the rth step at the kth frequency at the tth moment, denotes the ith frequency domain causal delay matrix of the rth step at the kth frequency at the tth moment, denotes the ith frequency domain topological causal matrix of the r-1th step at the kth frequency at the tth moment, q +1th frequency domain topological causal matrix, denotes the ith frequency domain causal delay matrix of the rth step at the kth frequency at the tth moment, q +1th frequency domain causal delay matrix, denotes the row vector of the ith row in the extraction matrix, i m denotes the quotient obtained by dividing the index i by 16, i q denotes the quotient obtained by dividing the index i by 16, i m denotes the remainder obtained by dividing the index i by 16, r denotes the number of steps of topological propagation, R denotes the assumed propagation step length, r = 1, 2,..., R, i denotes the node index number, i = 1, 2,..., 16 r ; at each frequency point k, the frequency domain delay perception value is calculated by weighted superposition of the delay matrix and the topological causal strength matrix of each step of topological propagation:

[0061]

[0062] wherein, v k,l (t) denotes the frequency domain delay perception value of the lth time step at the kth frequency at the tth moment, [·] mn denotes the element in the mth row and the nth column in the extraction matrix, exp(·) denotes the exponential function operation, l is the time step index, denotes the time sequence position corresponding to different delay steps, m denotes the node row index, n denotes the node column index, l = -L + 1,..., L - 1, m = 1, 2,..., 16, n = 1, 2,..., 16; according to the frequency domain delay perception result, the frequency domain connection strength matrix is constructed:

[0063] [M k (t)] gh = v k,g-h (t), -L + 1 ≤ g - h ≤ L - 1 (17)

[0064] wherein, M k (t) denotes the frequency domain connection strength matrix at the kth frequency at the tth moment, when the time step difference g - h is within the effective range [-L + 1, L - 1], when the time step difference connection strength matrix element is obtained by the delay perception value v k,g-h(t) = 0, otherwise, g denotes the connection strength matrix row index, h denotes the connection strength matrix column index, g = 1, 2, …, L, n = 1, 2, …, L; constructing a frequency domain topological organization attention, adjusting the connection strength between the query matrix and the key matrix through the frequency domain connection strength matrix, dynamically correcting the attention weight distribution, and making the attention allocation conform to the multi-path propagation relationship between the states of nodes; calculating the query matrix, the key matrix and the value matrix of the s th encoding module: Q s,k (t) = T s,k (t)W Q,s (t), K s (t) = T s,k (t)W K,s (t), V s (t) = T s,k (t)W V,s (t), Q s,k (t) denotes the query matrix of the s th encoding module under frequency k at time t, T s,k (t) denotes the input frequency domain feature token matrix under frequency k of the s th encoding module at time t, W Q,s (t) denotes the query weight matrix of the s th encoding module at time t, K s,k (t) denotes the key matrix under frequency k of the s th encoding module at time t, W K,s (t) denotes the key weight matrix of the s th encoding module, V s,k (t) denotes the value matrix under frequency k of the s th encoding module at time t, W V,s (t) denotes the value weight matrix of the s th encoding module at time t; calculating the frequency domain topological organization attention output matrix:

[0065] O s,k (t) = softmax(Q s,k (t)K s,k (t) T ⊙M k (t))V s,k (t) (18)

[0066] wherein, O s,k (t) denotes the frequency domain topological organization attention output matrix under frequency k of the s th encoding module at time t, softmax(·) denotes normalizing the matrix in the parentheses in the row dimension, so that the sum of each row element is 1; performing residual connection operation and layer normalization operation on the attention output matrix O s,k (t) under frequency k of the s th encoding module, to obtain the residual update output matrix R s,k (t) under frequency k of the s th encoding module; performing residual connection operation and layer normalization operation on the residual update output matrix R s,k(t) inputting the feedforward neural network to obtain the feedforward network output matrix of the s-th encoding module at frequency k The feedforward neural network structure comprises a first linear mapping layer with an input dimension of D and an output dimension of D, and a second linear mapping layer with an input dimension of D and an output dimension of D, and a nonlinear activation layer connected between the two, and the nonlinear activation function is a linear rectifier function; W 2,s (t) represents the first linear mapping weight matrix of the s-th encoding module, b 2,s (t) represents the first linear mapping bias vector of the s-th encoding module, W 3,s (t) represents the second linear mapping weight matrix of the s-th encoding module, b 3,s (t) represents the second linear mapping bias vector of the s-th encoding module; the feedforward network output matrix of the s-th encoding module at frequency k performing residual connection operation and layer normalization operation to obtain the frequency domain feature token matrix T output at frequency k of the s-th encoding module at time t s+1,k (t), represents the p-th row vector of the frequency domain feature token output matrix of the s-th encoding module at frequency k, p represents the index of the frequency domain feature token dimension, p = 1, 2,..., L;

[0067] Output layer: calculate the frequency domain feature prediction model output:

[0068]

[0069] wherein, represents the next time city sewage treatment supply chain product quality vector predicted by the model at time t, y1(t) represents the effluent total nitrogen concentration predicted by the model at time t, y2(t) represents the effluent total phosphorus concentration predicted by the model at time t, represents the row vector of the frequency domain feature token tensor unfolded by the output of the third encoding module, W4(t) represents the weight matrix of the output layer linear regression at time t, b4(t) represents the bias vector of the linear regression at time t;

[0070] 3. Training of the city sewage treatment supply chain product quality frequency domain topology feature prediction model:

[0071] ① Define the loss function of the model as:

[0072]

[0073] wherein, J(t) represents the loss of the model at time t, y(t-1) represents the actual next time urban sewage treatment supply chain product quality vector at t-1 time, y(t-1) = [x 1 (t), x 2 (t)] that is, the actual next time supply chain product quality at t-1 time is obtained by the actual data of total nitrogen concentration and total phosphorus of effluent collected at t time, which is used as a t time supervision signal to participate in model training;

[0074] ②Set the current training time as t, initialize the training round number τ = 1, and set the training iteration number as 20; initialize the weight parameters and bias parameters of the model, the weight parameters are randomly taken in the interval [-0.2, 0.2], and the bias parameter is taken as 0;

[0075] ③Calculate the prediction output of the model of the τth training at t-1 time by using formulas (1)-(7) Calculate the loss J of the model of the τth training at t time by using formula (8) τ (t), update the weight matrix and bias matrix by using the gradient descent method, and the calculation formula is:

[0076]

[0077] Wherein, W τ (t) represents the model weight matrix set of the τth training at t time, including: the encoding module query weight matrix, the encoding module key weight matrix, the encoding module value weight matrix, the encoding module first linear mapping weight matrix, the encoding module second linear mapping weight matrix and the output layer linear regression weight matrix, J represents the loss function J τ (t) of the τth training at t time, and the gradient of the model weight matrix set W τ (t) is W τ+1 (t) represents the model weight matrix set of the τ+1th training at t time, b τ (t) represents the model bias vector set of the τth training at t time, including: the encoding module first linear mapping bias vector, the encoding module second linear mapping bias vector and the output layer linear regression bias vector, J represents the loss function J τ (t) of the τth training at t time, and the gradient of the model bias vector set b τ (t) is b τ+1 (t) represents the model bias vector set of the τ+1th training at t time, τ represents the index of the training round number, τ = 1, 2, …, 20;

[0078] If the training round number τ is less than 20, τ is increased by 1, and the step 3 is continued; if the training round number τ is greater than or equal to 20, the model parameter training update is terminated, and the step 4 is entered to perform prediction;

[0079] 4. Predicting the product quality of the urban sewage treatment supply chain:

[0080] The trained product quality frequency domain topological feature prediction model of the urban sewage treatment supply chain is used, the influent flow, the external reflux flow, the influent ammonia nitrogen concentration, the anaerobic tank oxidation reduction potential, the anoxic zone import oxidation reduction potential, the anoxic zone nitrate nitrogen concentration, the anoxic zone mixed liquid suspended solid concentration, the internal reflux flow, the dissolved oxygen concentration of the front section of the aerobic zone, the dissolved oxygen concentration of the rear section of the aerobic zone, the first aeration quantity, the second aeration quantity, the third aeration quantity, the sludge discharge quantity, the effluent total nitrogen concentration and the effluent total phosphorus concentration collected at the time t are taken as the inputs of the model, and the model output vector at the time t is obtained according to the formula (12)-(19) The model output vector is taken as the predicted next time urban sewage treatment supply chain product quality at the time t, and the urban sewage treatment supply chain product quality includes the effluent total nitrogen concentration with the unit of mg / L and the effluent total phosphorus concentration with the unit of mg / L.

[0081] The effluent total nitrogen concentration and the effluent total phosphorus concentration prediction results of the method are shown in Figure 1 and Figure 2 The X axis is time with the unit of hour, and the Y axis is the effluent total nitrogen concentration prediction output and the effluent total phosphorus concentration prediction output with the unit of mg / L respectively; the solid line represents the actual measurement value, and the dotted line represents the frequency domain topological feature prediction output value. As shown in the figure, the prediction results of the method are basically consistent with the change trend of the actual measurement value in the whole period, and the future state of the urban sewage treatment supply chain product quality can be accurately reflected.

[0082] The prediction error results of the method are shown in Figure 3 and Figure 4 The X axis is time with the unit of hour, and the Y axis is the effluent total nitrogen concentration prediction error and the effluent total phosphorus concentration prediction error with the unit of mg / L respectively; the prediction error is calculated by the difference between the actual measurement value and the prediction value. As shown in the figure, the error fluctuation of the method is small in the whole prediction period, which shows that the method has good prediction stability and robustness in the whole period, and can effectively improve the accuracy and reliability of the urban sewage treatment supply chain product quality prediction.​

Claims

1. A municipal sewage treatment supply chain product quality frequency domain topological feature prediction method, characterized in that: Collecting urban sewage treatment operation data, establishing an urban sewage treatment supply chain product quality frequency domain topology characteristic prediction model, training an urban sewage treatment supply chain product quality frequency domain topology characteristic prediction model, and predicting urban sewage treatment supply chain product quality, comprising the following steps: (1) Collecting urban sewage treatment operation data: Taking urban wastewater treatment plants as the research object, the operation process of these plants includes influent transportation, external recirculation, internal recirculation, anaerobic reaction, anoxic reaction, aerobic reaction, sludge discharge, and effluent discharge. These stages are sequentially connected to form the operational units of the urban wastewater treatment supply chain. The treated effluent is considered the product of this supply chain, and its quality is determined by the effluent quality. The effluent quality is primarily reflected by the total nitrogen concentration and total phosphorus concentration, indicating the purification effect and effluent compliance level of the supply chain. Multi-source operational data reflecting flow rate, water quality, and reaction zone characteristics are collected. The influent flow rate x at time t is collected using an electromagnetic flowmeter installed on the outlet pipe of the influent pump station. 1 (t), in cubic meters per hour, is the external return flow rate x at time t, collected using an electromagnetic flowmeter installed on the outlet pipe of the external return pump. 2 (t), in cubic meters per hour, is the ammonia nitrogen concentration x in the influent at time t, collected using an online ammonia nitrogen analyzer installed on a branch pipe of the influent pipeline. 3 (t), in milligrams per liter, is the redox potential x of the anaerobic tank at time t, collected using a redox potential sensor installed inside the anaerobic tank. 4 (t), in millivolts, is the redox potential x at time t at the inlet of the anoxic zone, collected using a redox potential sensor installed at the inlet of the anoxic zone. 5 (t), in millivolts, represents the nitrate nitrogen concentration x in the anoxic zone at time t, collected using an online nitrate nitrogen analyzer installed in the anoxic zone reaction tank. 6 (t), in milligrams per liter, represents the concentration of suspended solids in the anoxic zone mixture at time t, measured using a suspended solids concentration meter installed in the anoxic zone reaction tank. 7 (t), in milligrams per liter, is the recirculation flow rate x at time t, collected using an electromagnetic flowmeter installed in the outlet pipe of the internal recirculation pump. 8 (t), in cubic meters per hour, is the dissolved oxygen concentration x in the aerobic zone at time t, collected using a dissolved oxygen sensor installed at the front of the aerobic zone. 9 (t), in milligrams per liter, is the dissolved oxygen concentration x in the latter part of the aerobic zone at time t, collected using a dissolved oxygen sensor installed in the latter part of the aerobic zone. 10 (t), in milligrams per liter, is the first aeration rate x at time t, collected using a thermal mass flow meter installed in the first branch of the aeration pipe. 11 (t), in cubic meters per hour, is the second aeration rate x at time t, collected using a thermal mass flow meter installed on the second branch of the aeration pipe. 12 (t), in cubic meters per hour, is the third aeration rate x at time t, collected using a thermal mass flow meter installed on the third branch of the aeration pipe. 13 (t), in cubic meters per hour, is the sludge discharge rate x at time t, collected using an electromagnetic flowmeter installed on the excess sludge discharge pipeline. 14 (t), unit: cubic meters / hour, the total nitrogen concentration x of the effluent at time t is collected by an online total nitrogen analyzer installed in the effluent pipeline 15 (t), unit: milligrams / liter, the total phosphorus concentration x of the effluent at time t is collected by an online total phosphorus analyzer installed in the effluent pipeline 16 (t), unit: milligrams / liter; x 1 (t), x 2 (t), x 3 (t), x 4 (t), x 5 (t), x 6 (t), x 7 (t), x 8 (t), x 9 (t), x 10 (t), x 11 (t), x 12 (t), x 13 (t), x 14 (t), x 15 (t), x 16 (t) as the input variable of the product quality prediction model of the municipal sewage treatment supply chain at time t, the total nitrogen concentration and the total phosphorus concentration of the effluent at the next time as the output variable of the model; t represents the current time step, N is the number of sample sequences, t = 1, 2, …, N; (2) Establishing an urban sewage treatment supply chain product quality frequency domain topology characteristic prediction model: An urban sewage treatment supply chain product quality frequency domain topology characteristic prediction model based on a frequency domain topology organization neural network is constructed, which is composed of an input layer, a frequency domain feature embedding layer, a frequency domain topology feature coding layer, and an output layer connection; Input layer: the collected data at time t and its previous L-1 historical time steps are arranged in time sequence to form the input variable vector, X n (t) represents the nth input variable vector, X n (t) = [x n (t-L+1), …, x n (t)], x n (t-L+1) represents the value of the nth input variable at time step t-L+1, L is the length of the input variable time window, n represents the index of the input variable, n = 1, 2, …, 16; the output matrix of the input layer is X(t) = [X 1 (t) T , …, X n (t) T , …, X 6 (t) T ] T , T represents the transpose of the matrix; Frequency domain feature embedding layer: map the input layer output matrix to the frequency domain through discrete Fourier transform, decompose X(t) into K=L / 2 frequency spectrum matrices; perform inverse Fourier transform on each frequency spectrum matrix to obtain the corresponding time domain feature matrix, and stack the time domain feature matrix in the frequency dimension to form the frequency domain feature token tensor T1(t), where T(t)=[T 1,1 (t),...,T 1,k (t),...,T 1,K (t)], T 1,k (t) represents the frequency domain feature token matrix output by the frequency domain feature embedding layer at the kth frequency, k=1,2,…,K; The frequency domain topology feature encoding layer is composed of 3 encoding modules stacked, s = 1, 2, 3, s represents the number of the s-th encoding module; the influence weight of the input layer output matrix from the historical time step to the current time step is calculated by using the least square estimation method, a multivariate autoregressive model is constructed to represent the time sequence dependence relationship between nodes; the autoregressive model is mapped to the frequency domain space based on the power spectral density function, and the frequency domain transfer function matrix is calculated at K discrete frequency points; the frequency domain causal topology matrix C k (t) represents the frequency domain causal topology matrix at frequency k at time t, the phase delay difference between adjacent frequency points is calculated based on the frequency phase expansion of the frequency domain transfer function matrix, and the frequency domain causal delay matrix is obtained, τ k (t) represents the frequency domain causal delay matrix at frequency k at time t; the single-step frequency domain topology causal matrix and the single-step frequency domain topology delay matrix are calculated: wherein, denotes the single-step frequency-domain topological causal matrix at time t and frequency k, denotes the single-step frequency-domain topological delay matrix at time t and frequency k, denotes the Hadamard product operator, 1 16×16 denotes a 16x16 all-ones matrix, I 16 denotes a 16-dimensional unit column vector, denotes a binarization function that filters valid topological connections; frequency-domain causal and causal-delay matrices under multi-step topological propagation are computed by recursion: in, This represents the i-th frequency domain topological causality matrix at the r-th step with frequency k at time t. This represents the i-th frequency domain causal delay matrix at frequency k at time t. Represents the i-th step at frequency k at time t (r-1). q +1 frequency domain topological causality matrix Represents the i-th step at frequency k at time t. q +1 frequency domain causal delay matrix Indicates the extraction of the i-th element from the matrix m The row vector of the row, i q This represents the quotient obtained by dividing index i by 16, where i m Let represent the remainder when index i is divided by 16, r represent the number of steps in the topology propagation, R represent the assumed propagation step size, r = 1, 2, ..., R, and i represent the node index number, i = 1, 2, ..., 16. r At each frequency point k, the frequency domain delay sensing value is calculated by weighting and superimposing the delay matrix and topological causality strength matrix of each step of topological propagation. wherein v k,l (t) denotes the frequency-domain delay-aware value of the l-th time step at frequency k at time t, [·] mn denotes the element in the m-th row and the n-th column of the extraction matrix, exp(·) denotes the exponential function operation, l is the time step index, denotes the time sequence position corresponding to different delay steps, m denotes the node row index, n denotes the node column index, l = -L + 1,..., L - 1, m = 1, 2,..., 16, n = 1, 2,..., 16; according to the frequency-domain delay-aware result, a frequency-domain connection strength matrix is constructed: [M k (t)] gh = v k,g-h (t), -L + 1 < g - h < L - 1 (6) wherein M k (t) represents the frequency domain connection strength matrix at time t and frequency k, when the time step difference g-h is in the effective range [-L+1, L-1], the connection strength matrix element is determined by the delay perception value v k,g-h (t) = 0, otherwise take zero, g represents the connection strength matrix row index, h represents the connection strength matrix column index, g = 1, 2,..., L, n = 1, 2,..., L; the frequency domain topological organization attention is constructed, the connection strength between the query matrix and the key matrix is adjusted by the frequency domain connection strength matrix, so as to dynamically correct the attention weight distribution, and make the attention distribution conform to the multi-path propagation relationship between the states of nodes; the query matrix, the key matrix and the value matrix of the s-th encoding module are calculated: Q s,k (t) = T s,k (t)W Q,s (t), K s (t) = T s,k (t)W K,s (t), V s (t) = T s,k (t)W V,s (t), Q s,k (t) represents the query matrix of the s-th encoding module at time t and frequency k, T s,k (t) represents the input frequency domain feature token matrix of the s-th encoding module at time t and frequency k, W Q,s (t) represents the query weight matrix of the s-th encoding module at time t, K s,k (t) represents the key matrix of the s-th encoding module at time t and frequency k, W K,s (t) represents the key weight matrix of the s-th encoding module, V s,k (t) represents the value matrix of the s-th encoding module at time t and frequency k, W V,s (t) represents the value weight matrix of the s-th encoding module at time t; the frequency domain topological organization attention output matrix is calculated: O s,k (t) = softmax(Q s,k (t)K s,k (t) T ⊙M k (t))V s,k (t) (7) wherein, O s,k (t) denotes the frequency-domain topological attention output matrix of the s-th encoding module at frequency k at time t, and softmax(·) denotes normalizing the matrix in the parentheses in the row dimension, so that the sum of each row element is 1; the attention output matrix O s,k (t) of the s-th encoding module at frequency k is subjected to residual connection operation and layer normalization operation to obtain the residual update output matrix R s,k (t) of the s-th encoding module at frequency k; the residual update output matrix R s,k (t) is input into the feedforward neural network to obtain the feedforward network output matrix The feedforward neural network structure comprises a first linear mapping layer with an input dimension of D and an output dimension of D, and a second linear mapping layer with an input dimension of D and an output dimension of D, and a nonlinear activation layer connected between the two, and the nonlinear activation function is a linear rectifier function; W 2,s (t) denotes the first linear mapping weight matrix of the s-th encoding module, and b 2,s (t) denotes the first linear mapping bias vector of the s-th encoding module, W 3,s (t) denotes the second linear mapping weight matrix of the s-th encoding module, and b 3,s (t) denotes the second linear mapping bias vector of the s-th encoding module; the feedforward network output matrix is subjected to residual connection operation and layer normalization operation to obtain the frequency-domain feature token matrix T s+1,k (t) output by the s-th encoding module at frequency k at time t, denotes the p-th row vector of the frequency-domain feature token output matrix of the s-th encoding module at frequency k, and p denotes the index of the frequency-domain feature token dimension, p = 1, 2, …, L. Output layer: calculate the frequency domain feature prediction model output: wherein, y1(t) represents the model-predicted effluent total nitrogen concentration at time t, and y2(t) represents the model-predicted effluent total phosphorus concentration at time t, y1(t) represents the model-predicted effluent total nitrogen concentration at time t, and y2(t) represents the model-predicted effluent total phosphorus concentration at time t, represents the row vector of the frequency domain feature token tensor unfolded by the 3rd encoding module output, W4(t) represents the weight matrix of the output layer linear regression at time t, and b4(t) represents the bias vector of the linear regression at time t; (3) Training an urban sewage treatment supply chain product quality frequency domain topology characteristic prediction model: ① Define the loss function of the model as: wherein J(t) represents the loss of the model at time t, represents the next time city sewage treatment supply chain product quality vector predicted by the model at time t-1, y(t-1) represents the actual next time city sewage treatment supply chain product quality vector at time t-1, y(t-1) = [x 1 (t), x 2 (t)], that is, the actual next time supply chain product quality at time t-1 is obtained from the actual data of the total nitrogen concentration and the total phosphorus of the effluent collected at time t, which is used as a supervision signal at time t to participate in model training; ② Set the current training time as t, initialize the training round number τ = 1, and set the training iteration number as 20; initialize the model weight parameters and bias parameters, the weight parameters are randomly taken in the interval [-0.2, 0.2], and the bias parameter is taken as 0; ③ The predicted output of the model trained for the τth time at time t-1 is calculated using formulas (1)-(7) The loss J of the model trained for the τth time at time t is calculated using formula (8) τ (t), the weight matrix and the bias matrix are updated using the gradient descent method, and the calculation formula is: wherein W τ (t) represents the model weight matrix set of the τth training at time t, including: the encoding module query weight matrix, the encoding module key weight matrix, the encoding module value weight matrix, the encoding module first linear mapping weight matrix, the encoding module second linear mapping weight matrix and the output layer linear regression weight matrix, J (t) represents the loss function J in the τth training at time t τ (t) represents the gradient of the model weight matrix set W τ (t) with respect to the model weight matrix set W τ+1 (t) represents the model weight matrix set of the τ+1th training at time t, b τ (t) represents the model bias vector set of the τth training at time t, including: the encoding module first linear mapping bias vector, the encoding module second linear mapping bias vector and the output layer linear regression bias vector, J (t) represents the loss function J in the τth training at time t τ (t) represents the gradient of the model bias vector set b τ (t) with respect to the model bias vector set b τ+1 (t) represents the model bias vector set of the τ+1th training at time t, τ represents the index of the training round, τ = 1, 2, …, 20; ④ If the training round number τ < 20, τ increases by 1, and step ③ is continued to train; if the training round number τ ≥ 20, the model parameter training update is terminated, and step (4) is entered to perform prediction; (4) Predicting urban sewage treatment supply chain product quality: The trained urban sewage treatment supply chain product quality frequency domain topological feature prediction model is used, and the influent flow, external reflux flow, influent ammonia nitrogen concentration, anaerobic tank oxidation reduction potential, anoxic zone import oxidation reduction potential, anoxic zone nitrate nitrogen concentration, anoxic zone mixed liquid suspended solid concentration, internal reflux flow, aerobic zone front section dissolved oxygen concentration, aerobic zone rear section dissolved oxygen concentration, first aeration quantity, second aeration quantity, third aeration quantity, sludge discharge quantity, effluent total nitrogen concentration and effluent total phosphorus concentration collected at time t are taken as inputs of the model, and the model output vector at time t is obtained according to formulas (1)-(8) The model output vector is taken as the predicted next time urban sewage treatment supply chain product quality at time t, and the urban sewage treatment supply chain product quality includes: effluent total nitrogen concentration, unit: mg / L, effluent total phosphorus concentration, unit: mg / L.

Images