Sewage treatment whole-process operation regulation and control calculation system based on deep learning

Abstract

The invention discloses a sewage treatment whole-process operation regulation and control calculation system based on deep learning, and relates to the technical field of intelligent control, and the system comprises a strategy optimization module which optimizes a future pollution trend prediction sequence and a regulation and control factor weight table by using a reinforcement learning strategy network in combination with a genetic algorithm optimizer, dynamically generating an aeration, dosing and backflow operation parameter combination aiming at a regulation target to form a dynamic regulation instruction set; the instruction execution module is used for issuing a dynamic regulation and control instruction set to the programmable logic controller by utilizing an industrial control interface, and collecting response information and real-time effluent quality data as execution feedback information; and the model updating module carries out error comparison on execution feedback information and a future pollution trend prediction sequence, constructs a weighted error sequence, and dynamically corrects deep network parameters through an online fine tuning strategy to form a prediction control model. According to the method, the generalization ability and robustness of the predictive control model are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control, and particularly relates to a sewage treatment full-process operation regulation and control computing system based on deep learning. BACKGROUND

[0002] With the acceleration of urbanization and the continuous enhancement of industrial activities, sewage treatment plays an increasingly important role in ensuring water environment safety and achieving sustainable development. At present, sewage treatment plants mostly use automatic control combined with online water quality monitoring to realize the operation regulation and control of key process units such as aeration, dosing and reflux. In recent years, intelligent control methods based on modeling and optimization have been introduced into the field of sewage treatment, including the use of model predictive control (MPC), fuzzy control and machine learning methods, to achieve a comprehensive balance of pollutant removal efficiency, energy consumption control and operation stability. In particular, the advantages of deep learning in time series modeling and nonlinear fitting make it gradually show strong application potential in sub-tasks such as water quality prediction, process anomaly detection and energy consumption modeling.

[0003] The existing technology still faces certain challenges in actual operation. Due to the high coupling, dynamicity and time-varying nature of the sewage treatment process, traditional control methods often rely on static models or empirical rules, and it is difficult to adapt to complex water quality disturbances and operation target switching in real time, especially in terms of multi-objective optimization regulation and control, lacking end-to-end dynamic response capability. Existing deep learning-based sewage treatment prediction methods mostly focus on a local task (such as water quality prediction or fault identification), lacking integrated modeling and collaborative optimization support for the whole process regulation chain. In the face of external sudden load, water quality fluctuations or variable operation parameter scenarios, it is difficult to achieve full-process collaborative perception, intelligent decision-making and online updating. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a deep learning-based sewage treatment full-process operation regulation and control computing system to solve the problems of dynamic optimization and self-learning regulation and control in the sewage treatment process.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The present application provides a deep learning-based sewage treatment full-process operation regulation and control computing system, which comprises a pretreatment module, which collects monitoring data and performs missing value repair and outlier rejection processing to construct a multivariate time series input matrix;

[0008] The feature extraction module sequentially inputs a multivariate time sequence input matrix into a convolutional neural network to extract short-term disturbance features, and then inputs the short-term disturbance features into a bidirectional long short-term memory network to extract long-term trend features, fuses the short-term disturbance features and the long-term trend features, and generates a future pollution trend prediction sequence and a regulation factor weight table through a multi-head time sequence attention mechanism to weight time periods and pollution factors;

[0009] The strategy optimization module optimizes the future pollution trend prediction sequence and the regulation factor weight table by using a reinforcement learning strategy network combined with a genetic algorithm optimizer, dynamically generates aeration, dosing and reflux operation parameter combinations for the regulation target, and forms a dynamic regulation instruction set.

[0010] The instruction execution module uses an industrial control interface to issue the dynamic regulation instruction set to a programmable logic controller, and simultaneously collects response information and real-time effluent water quality data as execution feedback information.

[0011] The model updating module compares the execution feedback information with the future pollution trend prediction sequence to construct a weighted error sequence, and dynamically corrects the deep network parameters through an online fine-tuning strategy to form a prediction control model.

[0012] As a preferred scheme of the sewage treatment full-process operation regulation and control computing system based on deep learning, the multivariate time sequence input matrix is constructed by the following steps,

[0013] The monitoring data is subjected to missing value filling and outlier removal processing to obtain cleaned monitoring data.

[0014] The cleaned monitoring data is arranged in time sequence and subjected to normalization processing to obtain standardized time sequence data.

[0015] The standardized time sequence data is divided into fixed-length time windows to form a multivariate time sequence input matrix.

[0016] As a preferred scheme of the sewage treatment full-process operation regulation and control computing system based on deep learning, the future pollution trend prediction sequence and the regulation factor weight table are generated by the following steps,

[0017] The multivariate time sequence input matrix is input into a convolutional neural network to capture the dynamic change trend of different variables in a local time window and extract a short-term disturbance feature map.

[0018] The short-term disturbance feature map is input into a bidirectional long short-term memory network to fuse the forward and backward context information in the time sequence, extract a long-term trend vector, and perform feature-level splicing and nonlinear fusion with the short-term disturbance feature map to obtain a unified time sequence representation.

[0019] The unified time sequence representation is input into a multi-head time sequence attention mechanism to focus on the correlation strength between time segments and pollution factors from different attention dimensions, and a weighted feature representation is generated.

[0020] The weighted feature representation is input into a fully connected prediction layer to generate a future pollution trend prediction sequence through nonlinear mapping, and the attention weights are analyzed to output a regulation factor weight table.

[0021] As a preferred scheme of the deep learning-based sewage treatment full-process operation regulation and control calculation system, the unified time sequence representation is input into a multi-head time sequence attention mechanism to focus on the correlation strength between time segments and pollution factors from different attention dimensions, and a weighted feature representation is generated, and the specific steps are as follows,

[0022] The unified time sequence representation is divided into multiple subspaces, and each attention head is input into the corresponding query, key and value vectors, and the attention weight matrix is generated through dot product operation;

[0023] The attention weight matrix is strengthened in the high weight area by using a sparse strategy, and the low weight time points and pollution factors are weakened to generate a sparse attention weight matrix and weight the corresponding value vectors to generate a weighted feature representation.

[0024] As a preferred scheme of the deep learning-based sewage treatment full-process operation regulation and control calculation system, the specific steps of forming the dynamic regulation instruction set are as follows,

[0025] The future pollution trend prediction sequence and the regulation factor weight table are input into a reinforcement learning strategy network to generate an initial strategy candidate set;

[0026] The initial strategy candidate set is iteratively optimized by using a genetic algorithm optimizer to select, cross and mutate, the optimal initial strategy is extracted as a feedback signal, the reinforcement learning strategy network parameters are updated, and an updated reinforcement learning strategy network is generated;

[0027] The updated reinforcement learning strategy network automatically generates an optimal operation parameter combination based on the comprehensive evaluation of the operation state and the regulation target, and integrates to form a dynamic regulation instruction set.

[0028] As a preferred scheme of the deep learning-based sewage treatment full-process operation regulation and control calculation system, the initial strategy candidate set is iteratively optimized by using a genetic algorithm optimizer to select, cross and mutate, the optimal initial strategy is extracted as a feedback signal, the reinforcement learning strategy network parameters are updated, and an updated reinforcement learning strategy network is generated, and the specific steps are as follows,

[0029] The initial strategy candidate set is scored by using a multi-objective fitness evaluation function to output a multi-dimensional fitness score vector.

[0030] The multi-dimensional fitness score vector is selected, crossed and mutated, high fitness individuals are preferentially retained and diversity disturbance is fused to generate a new generation of strategy population;

[0031] The new generation of strategy population is finely searched and clustered and reduced, the optimal strategy is extracted as a feedback signal, the reinforcement learning strategy network parameters are updated, and the updated reinforcement learning strategy network is generated.

[0032] As a preferred scheme of the sewage treatment full-process operation regulation and control calculation system based on deep learning, wherein: the response information and the real-time effluent water quality data are collected simultaneously as the execution feedback information, and the specific steps are as follows,

[0033] The dynamic regulation and control instruction set is converted into an industrial control interface compatible format, and is issued to the programmable logic controller through the industrial control interface to drive the aeration, dosing and reflux control actions;

[0034] The programmable logic controller collects the execution state and control response signal, and synchronously collects the real-time effluent water quality data to form joint execution feedback data;

[0035] The joint execution feedback data is preprocessed and verified, and the execution feedback information is output.

[0036] As a preferred scheme of the sewage treatment full-process operation regulation and control calculation system based on deep learning, wherein: the programmable logic controller collects the execution state and control response signal, and synchronously collects the real-time effluent water quality data to form joint execution feedback data, and the specific steps are as follows,

[0037] The programmable logic controller collects the execution state signal of aeration, dosing and reflux in real time, outputs the state data stream, and synchronously collects the control response signal to fuse and generate an execution action feature vector;

[0038] The real-time effluent water quality data is collected, and the time stamp and the execution action feature vector are synchronously matched to output a joint time series data set;

[0039] The joint time series data set is checked and corrected in real time by an anomaly detection algorithm, and the abnormal values are removed and the missing data is completed to output the joint execution feedback data.

[0040] As a preferred scheme of the sewage treatment full-process operation regulation and control calculation system based on deep learning, wherein: the actual pollutant concentration in the execution feedback information and the predicted value at the corresponding time point are aligned point by point, the absolute error, the relative error and the weighted error are calculated to generate a multi-dimensional error vector, and the multi-dimensional error vector is weighted and summarized according to the regulation factor weight table to output a weighted error sequence.

[0041] As a preferred scheme of the sewage treatment full-process operation regulation and control calculation system based on deep learning, the forming of the prediction control model comprises the following specific steps,

[0042] The weighted error sequence is analyzed in time sequence-factor dimension, the error high-sensitive area is extracted, and the reverse projection is performed to the characteristic activation map to identify the time slice and the high-sensitive parameter channel.

[0043] The time slice and the high-sensitive parameter channel are used to screen the parameter subset in the deep network, limit the fine-tuning range, reduce the calculation burden, focus on the influencing factors, construct the fine-tuning target function, and perform selective parameter updating to obtain the correction model structure.

[0044] The correction model structure is subjected to short-term back testing and stability verification, and the performance after improvement is evaluated as the prediction control model.

[0045] The beneficial effects of the present application are: by sequentially extracting short-term disturbance features and long-term trend features, and using multi-head time sequence attention mechanism to dynamically weight different time periods and pollution factors, the deep spatio-temporal correlation modeling and key feature enhancement of multi-dimensional time sequence data are realized, the accuracy and response sensitivity of pollution trend prediction are improved, the generalization ability and robustness of the prediction control model are significantly improved, a more scientific and reliable prediction basis is provided for sewage treatment operation regulation and control, and the dynamic regulation and control demand under complex environment is met. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Fig. 1 It is a schematic diagram of the sewage treatment full-process operation regulation and control calculation system based on deep learning in the present application.

[0048] Fig. 2 It is a flowchart of the multi-head time sequence attention mechanism processing in the present application.

[0049] Fig. 3 It is a flowchart of the genetic algorithm optimized reinforcement learning strategy network in the present application.

[0050] Fig. 4 It is a flowchart of the feedback information acquisition and processing in the present application. DETAILED DESCRIPTION

[0051] In order to make the above objectives, characteristics and advantages of the present application more apparent, concrete embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0052] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. The present application, however, can be practiced in a variety of ways other than those specifically described herein, and the scope of the present application is not limited to the specific embodiments described herein.

[0053] Secondly, the term "one embodiment" or "an embodiment" as used herein means that a particular implementation can include a particular feature, structure, or characteristic, but every embodiment can not necessarily include the particular feature, structure, or characteristic. Furthermore, the following terms can be used throughout the specification and claims: "include" means including but not limited to; "includes" means including but not limited to; "including" means including but not limited to; "comprise" means comprising but not limited to; "comprises" means comprising but not limited to; "comprised of" means comprising but not limited to; "comprising" means comprising but not limited to.

[0054] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides a deep learning-based sewage treatment full-process operation regulation and control computing system, comprising the following steps:

[0055] The preprocessing module collects monitoring data, and performs missing value repair and outlier rejection processing, and constructs a multivariate time series input matrix.

[0056] The monitoring data is subjected to missing value filling and outlier rejection processing to obtain cleaned monitoring data.

[0057] Specifically, when collecting monitoring data, online sensors and monitors are deployed according to key process sections in the sewage treatment process, and monitoring parameters include inflow and outflow, dissolved oxygen concentration, chemical oxygen demand, ammonia nitrogen, total phosphorus, pH value, temperature, stirring frequency, aeration time, dosing amount, and reflux ratio.

[0058] Each monitoring parameter is synchronously collected by an industrial automation data acquisition interface at a fixed sampling period (e.g., every 5 minutes), and a uniform format timestamp is attached; CRC cyclic redundancy check algorithm is used for integrity check to eliminate invalid data; the valid data is arranged in time sequence to construct an original monitoring data sequence; the original monitoring data sequence is aligned according to variable dimension and timestamp to form an initial monitoring data time series;

[0059] The missing value is processed by using a sliding window mean filling strategy. When the missing points are less than 3, for example, the mean values of the time points before and after are used for interpolation. When the missing points are more than 3, for example, the data at the corresponding time of the adjacent period is used to replace. The box plot method is used to determine the abnormal value. The values exceeding the range of [Q1-1.5*IQR, Q3+1.5*IQR] are replaced by the sliding median, and the cleaned monitoring data is output. Wherein, Q1 refers to the value located at the 25% position after a set of data is sorted from small to large, Q3 refers to the value located at the 75% position after sorting, and IQR represents the difference between Q3 and Q1.

[0060] The cleaned monitoring data is arranged in time sequence for normalization processing to obtain standardized time series data.

[0061] Specifically, after the cleaned monitoring data is arranged in time sequence, the maximum and minimum values of each monitoring parameter in the entire time sequence are counted respectively. The values of each monitoring parameter at each time point are linearly scaled according to the maximum and minimum values, so as to be mapped to the standard value interval between 0 and 1, and the normalization processing of all monitoring parameters is completed. The normalized result is used as the standardized time series data.

[0062] The standardized time series data is divided into time windows of fixed length to form a multivariate time series input matrix.

[0063] Specifically, the standardized time series data is arranged in time sequence, the sliding window length and step are unified, and continuous time segments are intercepted from the standardized time series data as a sample sequence in a sliding manner;

[0064] Each sample sequence contains multiple continuous time points, and each time point corresponds to the standardized value of all monitoring parameters. Multiple sample sequences are stacked in sequence to form a multivariate time series input matrix with uniform dimensions;

[0065] For example, if the time window length is set to 12 time points and the step is 1, 12 continuous time points are intercepted from the standardized time series data every 1 time point to form a sample sequence, and a complete multivariate time series input matrix is finally generated.

[0066] The feature extraction module inputs the multivariate time series input matrix into a convolutional neural network to extract short-term disturbance features, and then inputs the long-term trend features extracted by a bidirectional long short-term memory network. The short-term disturbance features and the long-term trend features are fused, and the time period and the pollution factor are weighted by a multi-head time attention mechanism to generate a future pollution trend prediction sequence and a control factor weight table.

[0067] The multivariate time series input matrix is input into a convolutional neural network to capture the dynamic change trend of different variables in a local time window, and a short-term disturbance feature map is extracted.

[0068] Specifically, the multivariate time series input matrix is organized into a tensor format acceptable to the convolutional neural network, where the feature dimension of the sample is taken as the number of channels, and the time step is taken as the sequence length, forming a three-dimensional input tensor with a shape of (batch size, feature number, time step), which is input into the convolutional neural network, and the convolution kernel size, step and padding mode are set to perform one-dimensional convolution operation on the multivariate time series input matrix.

[0069] The convolution kernel slides along the time dimension to extract the change characteristics of the monitoring data within the local time window; each convolution operation produces a local feature response, and a plurality of channel short-term disturbance feature maps are obtained through parallel operation of a plurality of convolution kernels.

[0070] The convolution operation is followed by batch normalization processing to stabilize the feature distribution, and a nonlinear activation function is applied to process the convolution output; in the example, the convolution kernel size is set to 3, the step is set to 1, and the number of channels is set to 64, and finally the output is a short-term disturbance feature map aligned with the original time dimension.

[0071] It should be noted that the specific steps of setting the convolution kernel size, step and padding mode are as follows: when processing the multivariate time series input matrix, the convolution kernel is set to, for example, 3 or 5 in the time dimension to extract local features of adjacent time steps, and is set to include all monitoring parameter channel numbers in the variable dimension to cover all variables; the step is set to, for example, 1 to maintain time continuity; and the padding mode is symmetric zero padding in the time dimension to make the output feature map length consistent with the input.

[0072] The short-term disturbance feature map is input into the bidirectional long short-term memory network to fuse the forward and backward context information in the time series, extract the long-term trend vector, and perform feature-level splicing and nonlinear fusion with the short-term disturbance feature map to obtain a unified time series representation.

[0073] Specifically, the short-term disturbance feature map output by the convolutional neural network is transposed in the dimension to convert it into a time series format of (batch size, time step, feature dimension); the time series is sequentially input into the forward and backward units of the bidirectional long short-term memory network to calculate the forward hidden state vector and the backward hidden state vector at each time step; the forward hidden state vector and the backward hidden state vector at each time step are spliced to form a long-term trend vector sequence corresponding to the time step;

[0074] The long-term trend vector sequence and the original short-term disturbance feature map are spliced at each time step in the feature dimension to obtain a fused feature sequence; the fused feature sequence is processed through a layer of nonlinear activation function (such as ReLU function) to complete the feature-level nonlinear fusion and generate a unified time series representation.

[0075] The unified time sequence representation is input into a multi-head time sequence attention mechanism to focus on the correlation strength between time slices and pollution factors from different attention dimensions, and a weighted feature representation is generated.

[0076] Furthermore, the unified time sequence representation is divided into multiple subspaces, and each attention head is input respectively, the corresponding query, key and value vectors are calculated, and the attention weight matrix is generated through dot product operation.

[0077] Specifically, the unified time sequence representation is equally divided into multiple subspaces along the feature dimension, and a corresponding number of attention heads are input respectively; each attention head maps the input subspace through the query weight matrix, the key weight matrix and the value weight matrix to obtain the query vector H, the key vector K and the value vector V; the Softmax function is used for normalization to obtain the attention weight matrix, and the expression is:

[0078]

[0079] Wherein, A represents the attention weight matrix, d k represents the dimension of the key vector, and W represents the length of the time sequence.

[0080] The sparse strategy is used to strengthen the high weight area of the attention weight matrix, weaken the low weight time points and pollution factors, generate a sparse attention weight matrix, and weight the corresponding value vector to generate a weighted feature representation.

[0081] Specifically, the sparse strategy is used to process the attention weight matrix, a sparsification threshold is set to screen the high weight area, the elements corresponding to the time points and pollution factors with weight values lower than the sparsification threshold are set to zero to generate a sparse attention weight matrix; the sparse attention weight matrix and the corresponding value vector are multiplied to calculate the weighted feature representation; each attention head is independently executed to generate the weighted feature representation.

[0082] It should be noted that when setting the sparsification threshold, the weight distribution in the attention weight matrix is counted, and an appropriate percentile is selected as the sparsification threshold, for example, the weight value of the 80th percentile in the attention weight matrix; all weight elements lower than the sparsification threshold are set to zero, and only the weight elements higher than or equal to the sparsification threshold are retained to form the sparse attention weight matrix; the time points and pollution factors with high weights are focused on, and the influence on the low weight area is weakened.

[0083] The weighted feature representation is input into the fully connected prediction layer to generate a future pollution trend prediction sequence through nonlinear mapping, and the attention weight is analyzed to output a regulatory factor weight table.

[0084] Specifically, a linear transformation is first performed on the weighted feature representation, and an intermediate output vector is calculated using an attention weight matrix and a bias vector; an activation function such as ReLU is applied to the intermediate output vector to complete the nonlinear mapping, and a future pollution trend prediction sequence is output.

[0085] At the same time, the attention weight matrix in the multi-head time sequence attention mechanism is analyzed, and a regulatory factor weight table is generated through weighted summation or normalization processing; the future pollution trend prediction sequence and the regulatory factor weight table are respectively used as inputs for subsequent regulation and optimization.

[0086] The strategy optimization module uses a reinforcement learning strategy network combined with a genetic algorithm optimizer to optimize the future pollution trend prediction sequence and the regulatory factor weight table, dynamically generates aeration, dosing, and reflux operation parameter combinations for the regulation target, and forms a dynamic regulation instruction set.

[0087] It should be noted that the combination of the reinforcement learning strategy network and the genetic algorithm optimizer enables joint optimization of the future pollution trend prediction sequence and the regulatory factor weight table, breaking through the limitations of traditional single optimization methods in generating regulation parameters. Through the complementary advantages of the strategy exploration ability of reinforcement learning and the global search advantage of genetic algorithms, aeration, dosing, and reflux operation parameter combinations that better meet the regulation target can be dynamically generated, improving the adaptability and optimization effect of the regulation instruction set, significantly enhancing the precision and real-time response capability of pollutant control, and promoting the development of intelligent regulation technology towards higher efficiency and intelligence.

[0088] The beneficial effects of combining the reinforcement learning strategy network with the genetic algorithm optimizer enable dynamic optimization of the future pollution trend prediction sequence and the regulatory factor weight table, which is significantly different from the single use of traditional optimization algorithms or static parameter adjustment methods in the prior art. Compared with the prior art, it can adapt to environmental changes in real time, automatically adjust aeration, dosing, and reflux operation parameter combinations, form a dynamic regulation instruction set, effectively improve the regulation precision and response speed, reduce the need for manual intervention, enhance the intelligence and self-adaptation capability, solve the technical problem that the regulation strategy in the prior art cannot balance the global optimum and real-time performance, and demonstrate significant innovation and practical value.

[0089] The future pollution trend prediction sequence and the regulatory factor weight table are input into the reinforcement learning strategy network to generate an initial strategy candidate set.

[0090] Specifically, the future pollution trend prediction sequence and the regulatory factor weight table are input as inputs, data format conversion and feature encoding are performed to meet the input requirements of the reinforcement learning strategy network;

[0091] The coded data is input into the reinforcement learning strategy network, and the strategy feature vector is extracted through the calculation and activation function processing of multiple neurons in the network; based on the strategy feature vector, the corresponding action probability distribution is generated by using the output layer of the strategy network, a plurality of strategy action candidates are sampled to form an initial strategy candidate set;

[0092] The initial strategy candidate set includes a plurality of combinations of aeration, dosing and reflux operation parameters for different control targets.

[0093] The initial strategy candidate set is iteratively optimized by selection, crossover and mutation operations of the genetic algorithm optimizer, and the optimal initial strategy is extracted as a feedback signal to update the reinforcement learning strategy network parameters and generate an updated reinforcement learning strategy network.

[0094] Further, the initial strategy candidate set is scored for comprehensive performance by using a multi-objective fitness evaluation function, and a multi-dimensional fitness score vector is output.

[0095] Specifically, a plurality of performance indicators, such as aeration energy consumption, effluent water quality indicators and operation stability, are calculated for each candidate strategy as evaluation dimensions; the plurality of performance indicators are normalized according to weights to eliminate dimensional differences; the normalized values of each indicator are weighted and summed according to the multi-objective fitness evaluation function to obtain a multi-dimensional fitness score vector of the corresponding strategy; the steps are repeated to calculate the multi-dimensional fitness score vectors of all candidate strategies.

[0096] The multi-dimensional fitness score vector is selected, crossed and mutated to preferentially retain high fitness individuals and fuse diversity disturbance to generate a new generation of strategy population.

[0097] Specifically, when performing genetic operations on the multi-dimensional fitness score vector, a roulette wheel selection method is used to select a plurality of high fitness individuals into the next generation according to the fitness values, ensuring the retention of excellent strategies.

[0098] Randomly pair the selected individuals in pairs, exchange chromosome fragments based on single-point or multi-point crossover strategies, and generate new individuals to enhance strategy diversity.

[0099] According to the mutation probability, the strategy parameters of the crossed individuals are randomly adjusted at certain gene sites to produce slight disturbance; the selection, crossover and mutation process is repeated iteratively to form a new generation of strategy population containing high-quality and diverse strategies.

[0100] The new generation of strategy population is finely searched and clustered to extract the optimal strategy as a feedback signal to update the reinforcement learning strategy network parameters and generate an updated reinforcement learning strategy network.

[0101] Specifically, for the new generation of strategy groups, the performance indicators of each strategy are calculated in turn, and fine search is performed to obtain accurate performance evaluation results; a clustering algorithm is used to classify strategies with similar performance indicators into the same group, and the characteristic distribution of the strategies in each group is analyzed to determine the similarity between different strategy categories, and a representative strategy is selected in each cluster as a candidate; based on the performance evaluation and clustering results, the strategy with the best overall performance is selected as the feedback signal; the selected feedback signal is input into the parameter updating mechanism of the reinforcement learning strategy network to adjust the network weights and biases, and the parameter updating of the reinforcement learning strategy network is completed, and finally the updated reinforcement learning strategy network is generated.

[0102] The updated reinforcement learning strategy network automatically generates the optimal combination of operating parameters based on the comprehensive evaluation of the operating state and the control target, and integrates to form a dynamic control instruction set.

[0103] Specifically, according to the updated reinforcement learning strategy network, the current control target and the pre-processed monitoring data are input, and a set of continuous candidate values of aeration, dosing and reflux operating parameters are output by executing forward propagation calculation of the network; the continuous candidate values are discretized according to the pre-set operating parameter range and step size to generate a limited number of executable parameter value points; among the discretized parameter value points, the combinations of aeration, dosing and reflux operating parameters that meet the constraints of the sewage treatment process are selected; the selected operating parameter combinations are encoded according to the format specified by the industrial control protocol, and a dynamic control instruction set is integrated to generate a dynamic control instruction set, which contains specific aeration time, dosing dose and reflux ratio values.

[0104] For example, the aeration time range is 0 to 60 minutes, the step size is 5 minutes; the dosing dose range is 0 to 100 mg / L, the step size is 10 mg / L; the reflux ratio range is 50% to 200%, the step size is 10%.

[0105] It should be noted that the specific steps of the pre-set operating parameter range and step size are as follows: according to the sewage treatment process specification and performance, the minimum and maximum ranges of each operating parameter are determined, for example, the aeration time is set to 10 minutes to 60 minutes, the dosing dose is set to 0 to 50 ml, and the reflux ratio is set to 0.1 to 1.0; according to the control accuracy requirement, the step size of each operating parameter is set, for example, the aeration time step size is 5 minutes, the dosing dose step size is 1 ml, and the reflux ratio step size is 0.1; the discrete value set divided by the step size in the operating parameter range is formed into an operating parameter candidate set for subsequent parameter combination screening.

[0106] The instruction execution module uses the industrial control interface to issue the dynamic control instruction set to the programmable logic controller, and simultaneously collects response information and real-time effluent water quality data as execution feedback information.

[0107] The dynamic control instruction set is converted into an industrial control interface compatible format, and is sent to the programmable logic controller through the industrial control interface to drive the aeration, dosing and reflux control actions.

[0108] Specifically, the aeration, dosing and reflux operation parameters are encoded and packaged according to the communication protocol and data format required by the industrial control interface, for example, using Modbus or OPC protocol, each operation parameter is converted into a corresponding register address and value; a complete control message is generated, including instruction type, target address and check code; the control message is sent to the programmable logic controller through the industrial control interface communication channel; the programmable logic controller receives and parses the control message, and drives the corresponding aeration device, dosing device and reflux pump to execute the control action according to the instruction parameters.

[0109] The programmable logic controller collects the execution state and control response signals, and synchronously collects the real-time effluent water quality data to form joint execution feedback data.

[0110] Further, the programmable logic controller collects the execution state signals of aeration, dosing and reflux in real time, outputs the state data stream, and synchronously collects the control response signals to generate an execution action feature vector.

[0111] Specifically, the programmable logic controller collects the execution state signals of aeration, dosing and reflux in real time, including current value, frequency, running time and start-stop state, etc., acquires the state quantity through the input channel and generates a state data stream at a fixed period;

[0112] At the same time, the control response signals are collected, including the response deviation of the actual execution parameters and the instruction parameters in the dynamic control instruction set, such as the difference between the set frequency and the actual frequency; after aligning the state data stream and the control response signals according to the time stamp, the execution action feature vector is generated by fusion through feature encoding;

[0113] For example, different state quantities and response differences are combined into a one-dimensional feature array through splicing operation.

[0114] The real-time effluent water quality data is collected, and the joint time series data set is output by synchronously matching the time stamp and the execution action feature vector.

[0115] Specifically, the real-time effluent water quality data is collected through the water quality online monitoring instrument, including ammonia nitrogen concentration, total phosphorus concentration, total nitrogen concentration, chemical oxygen demand and dissolved oxygen concentration, etc., and the water quality raw data stream with time stamp is generated according to the sampling period;

[0116] Each group of time stamps in the effluent water quality data is matched one by one with the time stamps in the execution action feature vector, and the time series is aligned by using nearest neighbor matching or linear interpolation matching method.

[0117] After matching is completed, the water quality data at the corresponding time point is spliced with the execution action feature vector to generate a joint time sequence sample; the whole cycle data matching is completed by repeating the steps, and a joint time sequence dataset is output.

[0118] The joint time sequence dataset is checked and corrected in real time by an anomaly detection algorithm, and abnormal values are eliminated and missing data is completed, to output joint execution feedback data.

[0119] Specifically, an anomaly detection algorithm is performed on each field in the joint time sequence dataset, and an algorithm such as a local outlier factor algorithm or a sliding window Z-score method is used to detect abnormal values in ammonia nitrogen concentration, total phosphorus concentration, total nitrogen concentration, chemical oxygen demand, dissolved oxygen concentration, and aeration frequency, dosing amount, and reflux flow in the execution action feature vector, and observation points exceeding a set joint time sequence data anomaly identification threshold are marked as abnormal values.

[0120] The data points marked as abnormal are eliminated, and time series interpolation algorithms are used to complete the original missing positions at the eliminated positions, such as linear interpolation or sliding average method based on historical window for data reconstruction; after the whole field data cleaning is completed, the corrected joint time sequence dataset is uniformly output as joint execution feedback data.

[0121] It should be noted that the specific steps of setting the joint time sequence data anomaly identification threshold are as follows: reading the joint time sequence dataset, statistically analyzing each monitoring field such as ammonia nitrogen concentration, total phosphorus concentration, and dissolved oxygen concentration, calculating the historical mean and standard deviation of each field, setting the joint time sequence data anomaly identification threshold as the mean plus or minus three times the standard deviation, forming the upper and lower limits of the field, for example, when the mean of the ammonia nitrogen concentration field is 8 mg / L and the standard deviation is 1 mg / L, the joint time sequence data anomaly identification threshold is set to 5 mg / L to 11 mg / L, and subsequent observation data exceeding the range will be marked as abnormal values.

[0122] The joint execution feedback data is preprocessed and checked, and execution feedback information is output.

[0123] Specifically, missing value detection and filling processing are performed on each field in the joint execution feedback data, and a small amount of intermittent missing values are completed using linear interpolation method;

[0124] Secondly, the numerical range is checked, and according to the physical effective range of the field, for example, the range of the dissolved oxygen concentration field is set to 0 mg / L to 20 mg / L, and illegal values exceeding the range are eliminated;

[0125] Then, the time continuity is checked, and the difference operation is performed on the time stamp sequence to ensure that the time interval is within the allowed range, for example, the difference should not exceed 10 minutes, and if it exceeds, it is marked as abnormal;

[0126] Finally, the field unit is converted uniformly, such as converting milligrams per liter (mg / L) into a standard unit format, and the output is a feedback information with complete structure, numerical compliance, and continuous time series.

[0127] The model updating module compares the execution feedback information with the future pollution trend prediction sequence to build a weighted error sequence, and dynamically corrects the deep network parameters through an online fine-tuning strategy to form a prediction control model.

[0128] It should be noted that the innovation lies in the real-time fusion of execution feedback information and future pollution trend prediction sequence to build a weighted error sequence, realize dynamic quantization and multi-dimensional weighted processing of errors; based on the weighted error sequence, the deep network parameters are dynamically corrected in real time through an online fine-tuning strategy, breaking through the traditional static training mode, realizing adaptive updating and continuous optimization of the prediction control model, improving the response ability and prediction accuracy of the prediction control model to environmental changes and operating conditions, thereby effectively improving the refinement and real-time performance of pollutant regulation.

[0129] The beneficial effects are that by building a weighted error sequence of execution feedback information and future pollution trend prediction sequence, and dynamically correcting deep network parameters through an online fine-tuning strategy, the defects of fixed prediction model parameters in the prior art that are difficult to adapt to environmental dynamic changes are overcome. Compared with the traditional static training method, the prediction control model parameters are adjusted in real time, significantly improving the accuracy and robustness of the prediction control; at the same time, through multi-dimensional weighted error quantization, the accuracy of identifying and responding to pollution factors and time segments is enhanced, the pertinence and effectiveness of the control strategy are enhanced, and the real-time performance and refinement of pollutant concentration control are significantly improved, having better engineering application value and promotion potential.

[0130] The actual pollutant concentration in the execution feedback information and the predicted value at the corresponding time point are aligned point by point, the absolute error, relative error and weighted error are calculated, a multi-dimensional error vector is generated, the multi-dimensional error vector is weighted and summarized according to the control factor weight table, and a weighted error sequence is output.

[0131] Specifically, the actual pollutant concentration sequence and the predicted pollutant concentration sequence are matched point by point according to the time stamp to ensure one-to-one correspondence of the time points; the absolute error sequence is calculated, and the expression is:

[0132] E abs (t)=|C actual (t)-C predicted (t)|;

[0133] Wherein, E abs (t) represents the absolute error at time point t, C actual (t) represents the actual pollutant concentration at time point t in the execution feedback information, and Cpredicted (t) represents the predicted pollutant concentration at the corresponding time point, t represents a specific time point in the time sequence;

[0134] The relative error sequence is calculated, and the expression is:

[0135]

[0136] wherein E rel (t) represents the relative error at time point t, and ∈ represents a very small positive number to prevent division by zero;

[0137] Based on the corresponding pollutant weight w i in the regulatory factor weight table, the weighted error sequence is calculated, and the expression is:

[0138]

[0139] wherein E weighted (t) represents the weighted error at time point t, w i represents the weight coefficient corresponding to the i-th pollutant in the regulatory factor weight table, i represents the number of the pollutant, and n represents the total number of pollutant types, E abs,i (t) represents the absolute error of the i-th pollutant at time point t;

[0140] Finally, all time points t are traversed, and the weighted error values are arranged in time sequence to form a weighted error sequence.

[0141] The weighted error sequence is analyzed in time-factor dimension, the error high-sensitive area is extracted, and is reversely projected to the feature activation map to identify the time segment and the high-sensitive parameter channel.

[0142] Specifically, the weighted error sequence is constructed into a two-dimensional error matrix according to the time dimension and the regulatory factor dimension, the time dimension corresponds to each time point in the sequence, and the regulatory factor dimension corresponds to each pollutant or parameter channel;

[0143] The error matrix is segmented and counted in the time dimension by using the sliding window method, the error sum or mean value in each time window is calculated, and the time segment with an error value higher than a set error threshold is screened out, which is recorded as the error high-sensitive area;

[0144] The time window index corresponding to the error high-sensitive area is mapped back to the feature activation map, the weight of each parameter channel is evaluated in combination with the activation intensity, the high-sensitive parameter channel with a greater contribution in the error high-sensitive area is identified, and the error high-sensitive time segment range and the corresponding high-sensitive parameter channel are output.

[0145] It should also be noted that the specific steps of setting the error threshold: collect historical weighted error sequence samples, count the error distribution characteristics, calculate the mean and standard deviation of the error; according to the error statistics, select the appropriate confidence interval, for example, set the error threshold as the mean plus twice the standard deviation as the error threshold; adjust the error threshold value according to the actual control requirements and tolerance range, ensure that the error threshold can accurately identify the area of abnormally high error, and not be too sensitive to cause false positives; the determined error threshold is applied to the screening process of the error high sensitive area to determine whether the error exceeds the allowed fluctuation interval calculated by the mean and standard deviation.

[0146] Using time segments and high sensitive parameter channels, screening parameter subsets in deep networks, limiting fine-tuning range, reducing computational burden and focusing on influencing factors, constructing fine-tuning objective function, and performing selective parameter update to obtain corrected model structure.

[0147] Specifically, according to the identified time segments and high sensitive parameter channels, the corresponding feature regions are located from the unified time sequence representation, and the parameter range to be fine-tuned is determined; the fine-tuning objective function is constructed, and the weighted error sequence of the screened error high sensitive area is taken as the optimization object, and the influence of parameter adjustment on error is quantified combined with the control factor weight table;

[0148] The selected parameter subsets are selectively updated using existing technologies such as gradient descent, limiting parameter adjustment within a limited range to reduce computational burden and focus on main influencing factors;

[0149] After completing parameter fine-tuning, the model structure is updated to obtain the corrected model structure, ensuring that the correction process is efficient and highly targeted.

[0150] The corrected model structure is subjected to short-term backtesting and stability verification, and the performance improvement is evaluated as a predictive control model.

[0151] Specifically, recent historical joint execution feedback data is selected as the backtesting sample, which is input into the corrected model structure to generate the corresponding future pollution trend prediction sequence; the future pollution trend prediction sequence is compared with the actual monitoring value at the corresponding time point in the backtesting sample point by point, and the absolute error is calculated using the point-by-point difference method, and the relative error is determined by the ratio of the absolute error to the actual monitoring value, expressed in percentage form, to evaluate the prediction accuracy;

[0152] Check if the trend change of the prediction result on the pollution factor is consistent with the actual one, verify the trend consistency; analyze the distribution characteristics of the prediction error in the time dimension, determine whether there are periodic fluctuations or mutation points, and confirm the stability of the corrected model output; if the backtesting result shows that the prediction accuracy is improved, the trend consistency is strong, and the error fluctuation is within the allowed range, it is determined that the corrected model structure meets the preset performance standard, and is confirmed as a predictive control model.

[0153] For example, the preset performance standard is that the average relative error of COD prediction is less than 15%, the root mean square error of NH3-N prediction is less than 2.0 mg / L, and the error standard deviation is not more than 1.5 mg / L for 72 consecutive hours. The confirmed prediction control model is used for subsequent sewage treatment whole process operation regulation and control calculation, and the specific application mode is as follows: in each regulation and control period, the monitoring data collected in real time and processed by the preprocessing module are input into the prediction control model to generate the pollution trend prediction sequence and the regulation factor weight table in the future period, which are output to the strategy optimization module as the output of the feature extraction module to participate in the generation of the dynamic regulation and control instruction set, so as to realize the closed-loop optimization regulation and control of the aeration, dosing and reflux operation parameters.

[0154] In summary, the present application realizes the deep space-time correlation modeling and key feature enhancement of multi-dimensional time series data by sequentially extracting short-term disturbance features and long-term trend features and dynamically weighting different time periods and pollution factors by using a multi-head time series attention mechanism, improves the accuracy and response sensitivity of pollution trend prediction, significantly improves the generalization ability and robustness of the prediction control model, provides a more scientific and reliable prediction basis for sewage treatment operation regulation and control, and meets the dynamic regulation and control demand under complex environment.

[0155] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based computational system for the entire process of wastewater treatment operation and control, characterized in that: include, The preprocessing module collects monitoring data, performs missing value repair and outlier removal, and constructs a multivariate time series input matrix. The feature extraction module sequentially inputs the multivariate time series input matrix into a convolutional neural network to extract short-term perturbation features, and then inputs it into a bidirectional long short-term memory network to extract long-term trend features. The short-term perturbation features and long-term trend features are fused together, and the time period and pollution factors are weighted through a multi-head time series attention mechanism to generate a future pollution trend prediction sequence and a regulatory factor weight table. The strategy optimization module uses a reinforcement learning strategy network combined with a genetic algorithm optimizer to optimize the future pollution trend prediction sequence and the weight table of regulatory factors, and dynamically generates a combination of aeration, dosing and reflux operation parameters for the regulatory target, forming a dynamic regulation instruction set. The instruction execution module uses the industrial control interface to send the dynamic control instruction set to the programmable logic controller, and at the same time collects response information and real-time effluent water quality data as execution feedback information. The model update module compares the execution feedback information with the future pollution trend prediction sequence to construct a weighted error sequence, and dynamically corrects the deep network parameters through an online fine-tuning strategy to form a predictive control model.

2. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 1, characterized in that: The specific steps for constructing the multivariate time-series input matrix are as follows. The monitoring data is processed by filling in missing values ​​and removing outliers to obtain cleaned monitoring data; The cleaned monitoring data were arranged in chronological order and then normalized to obtain standardized time-series data. Standardized time series data are divided into fixed-length time windows to form a multivariate time series input matrix.

3. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 1, characterized in that: The specific steps for generating the future pollution trend prediction sequence and the weight table of regulatory factors are as follows. By inputting a multivariate time series input matrix into a convolutional neural network, the dynamic change trends of different variables within a local time window are captured, and short-term perturbation feature maps are extracted. The short-term perturbation feature map is input into a bidirectional long short-term memory network, which integrates forward and backward contextual information from the time series to extract the long-term trend vector. This vector is then combined with the short-term perturbation feature map through feature-level concatenation and nonlinear fusion to obtain a unified time series representation. The unified temporal representation is input into the multi-head temporal attention mechanism to focus on the correlation strength between time segments and pollutants from different attention dimensions, and generate a weighted feature representation. The weighted feature representation is input into the fully connected prediction layer, and a future pollution trend prediction sequence is generated through nonlinear mapping. At the same time, the attention weights are analyzed and the regulatory factor weight table is output.

4. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 3, characterized in that: The unified temporal representation is input into a multi-head temporal attention mechanism, which focuses on the correlation strength between time segments and contaminants from different attention dimensions to generate a weighted feature representation. The specific steps are as follows. The unified temporal representation is divided into multiple subspaces, each of which is input into an attention head. The corresponding query, key, and value vectors are calculated, and the attention weight matrix is ​​generated by dot product operation. By employing a sparsity strategy, the attention weight matrix is ​​strengthened in high-weight regions and weakened in low-weight time points and contamination factors, generating a sparse attention weight matrix and weighting the corresponding value vectors to produce a weighted feature representation.

5. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 1, characterized in that: The specific steps for forming the dynamic control instruction set are as follows: Input the future pollution trend prediction sequence and the weight table of regulatory factors into the reinforcement learning policy network to generate an initial policy candidate set; Using a genetic algorithm optimizer, the initial policy candidate set is iteratively optimized by selection, crossover and mutation operations. The optimal initial policy is extracted and used as a feedback signal to update the parameters of the reinforcement learning policy network and generate the updated reinforcement learning policy network. The updated reinforcement learning policy network automatically generates the optimal combination of operating parameters based on a comprehensive evaluation of the operating state and the control objective, and integrates them into a dynamic control instruction set.

6. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 5, characterized in that: The genetic algorithm optimizer iteratively optimizes the initial policy candidate set through selection, crossover, and mutation operations, extracts the optimal initial policy as a feedback signal, updates the reinforcement learning policy network parameters, and generates the updated reinforcement learning policy network. The specific steps are as follows. Using a multi-objective fitness evaluation function, the initial policy candidate set is comprehensively scored, and a multi-dimensional fitness score vector is output. By performing selection, crossover, and mutation operations on the multidimensional fitness score vector, high-fitness individuals are retained first and diversity perturbations are incorporated to generate a new generation of strategy populations. A refined search and cluster reduction of the new generation of policy families is performed to extract the optimal policy as a feedback signal, update the parameters of the reinforcement learning policy network, and generate the updated reinforcement learning policy network.

7. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 1, characterized in that: The simultaneous collection of response information and real-time effluent water quality data as execution feedback information involves the following specific steps. The dynamic control instruction set is converted into an industrial control interface compatible format and sent to the programmable logic controller through the industrial control interface to drive aeration, dosing and reflux control actions. The programmable logic controller collects execution status and control response signals, and simultaneously collects real-time effluent water quality data to form joint execution feedback data; The joint execution feedback data is preprocessed and verified, and the execution feedback information is output.

8. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 7, characterized in that: The programmable logic controller (PLC) collects execution status and control response signals, and simultaneously collects real-time effluent water quality data to form joint execution feedback data. The specific steps are as follows: The programmable logic controller (PLC) collects the execution status signals of aeration, dosing, and reflux in real time, outputs the status data stream, and simultaneously collects the control response signals, which are then fused to generate the execution action feature vector. Collect real-time effluent water quality data, synchronize and match it with the feature vector of the executed action according to the timestamp, and output a joint time series dataset; An anomaly detection algorithm is used to perform real-time verification and correction on the joint time-series dataset, removing outliers and filling in missing data, and outputting joint execution feedback data.

9. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 1, characterized in that: Align the actual pollutant concentrations in the execution feedback information with the predicted values ​​at the corresponding time points point by point, calculate the absolute error, relative error, and weighted error, generate a multidimensional error vector, and sum the weighted values ​​of the multidimensional error vector according to the control factor weight table to output the weighted error sequence.

10. The deep learning-based wastewater treatment process operation and control calculation system as described in claim 1, characterized in that: The specific steps for forming the predictive control model are as follows. The weighted error sequence is analyzed in time-series and factor dimensions to extract the high-sensitivity region of error, and then back-projected onto the feature activation map to identify time segments and high-sensitivity parameter channels. By utilizing time segments and highly sensitive parameter channels, a subset of parameters in deep networks is selected to limit the fine-tuning range, reduce computational burden, focus on influencing factors, construct a fine-tuning objective function, and perform selective parameter updates to obtain the corrected model structure. The calibration model structure was backtested and its stability verified in a short period of time. The improved performance was then evaluated and used as a predictive control model.

Images