A sewage treatment operation energy consumption optimization method and system based on deep learning

Abstract

The application discloses a sewage treatment operation energy consumption optimization method and system based on deep learning, which comprises the following steps: collecting multi-source operation data, preprocessing, forming a multi-dimensional operation state feature matrix; inputting an improved MICN model, performing multi-scale time series modeling, outputting water quality prediction values and energy consumption prediction values; performing controlled disturbance calculation, calculating the water quality improvement response of unit energy consumption, and performing energy consumption-water quality reverse contribution decomposition; constructing a water quality response reachable domain, performing consistency determination, and determining a feasible control interval; under the constraints of energy consumption structure and control interval, the optimal control parameters are obtained through collaborative optimization, and energy consumption inertia constraint is performed; steady-state boundary verification is performed, and if the standard constraint is met, control is performed, and the feature matrix is updated for closed-loop optimization. Through the improved MICN model prediction, reverse contribution decomposition and reachable domain constraint optimization, the application realizes the minimum energy consumption control of sewage treatment under stable standard.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method and system for optimizing energy consumption in wastewater treatment operations based on deep learning. Background Technology

[0002] Wastewater treatment processes are typical multivariate coupled, highly nonlinear, and non-stationary systems, with energy consumption primarily concentrated in key units such as aeration, recirculation, and chemical dosing systems. To reduce operating costs and ensure stable effluent quality, existing technologies typically employ energy optimization strategies based on empirical rules, mechanistic models, or conventional data-driven methods. For example, historical data regression analysis, single-time-series prediction models, or simple optimization algorithms are used to adjust aeration rates, recirculation ratios, and chemical dosages. While these methods are effective under steady-state or slightly fluctuating conditions, in actual wastewater treatment processes, the significant time-varying and stochastic nature of influent quality, quantity, and load makes it difficult for traditional methods to accurately characterize the complex temporal coupling relationships between operating parameters, energy consumption, and effluent quality.

[0003] With the development of deep learning technology, some existing technologies have begun to incorporate neural network models to predict effluent quality or energy consumption, and to carry out optimization control based on this prediction. However, existing data-driven methods mostly focus on the forward prediction process, that is, directly predicting future water quality or energy consumption based on historical operating data, and then optimizing overall parameters. They lack the ability to finely analyze the actual effect of each component of energy consumption on water quality improvement, resulting in redundant energy input even when meeting effluent standards. Existing optimization methods usually do not combine mechanistic constraints such as sludge age, hydraulic retention time, and biochemical reaction rate to construct an achievable operating range, which easily produces optimization results that deviate from the actual biochemical reaction capacity. Under load fluctuations or shock conditions, this can easily cause frequent fluctuations in control parameters, affecting operational stability.

[0004] Furthermore, existing technologies, when performing optimized control, often directly send the optimization results to the control system without fully considering factors such as the response characteristics of aeration equipment, the hysteresis characteristics of biochemical reactions, and operational inertia. This can easily lead to sudden changes in control parameters, causing fluctuations in dissolved oxygen and instability in sludge conditions, thereby affecting the safety of effluent quality. Under conditions of strong fluctuations in influent quality and coupling of multiple operating conditions, existing wastewater treatment operation energy consumption optimization methods generally suffer from technical defects such as difficulty in identifying the actual effective energy consumption contribution of each control unit, lack of consistency in optimization results based on mechanisms, and insufficient execution stability.

[0005] Therefore, how to provide a method and system for optimizing the energy consumption of wastewater treatment operations based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method and system for optimizing energy consumption in wastewater treatment operations. This invention constructs an improved MICN model to perform multi-scale time-series modeling of multi-source operational data, and combines energy consumption-water quality inverse contribution decomposition, water quality reaction reachability domain determination, and energy consumption response inertia constraint processing to achieve coordinated optimization control of aeration volume, reflux ratio, and reagent dosage. This allows for accurate identification of the actual role of each energy consumption component in improving effluent quality. Under the premise of ensuring stable effluent compliance, an equivalent minimum necessary energy consumption structure is constructed and a closed-loop optimization and control mechanism is formed. This method has the advantages of strong adaptability to non-stationary operating conditions, high energy consumption optimization accuracy, good mechanism consistency, and high operational stability.

[0007] According to an embodiment of the present invention, a wastewater treatment operation energy consumption optimization method based on deep learning includes:

[0008] Collect multi-source operational data from the wastewater treatment system, preprocess the multi-source operational data, and form a multi-dimensional operational status feature matrix;

[0009] The multi-dimensional operating state feature matrix is ​​input into the improved MICN model to perform multi-scale time series modeling of control variables, sub-item energy consumption and process state parameters, and output the predicted values ​​of effluent water quality and sub-item energy consumption in the corresponding prediction period.

[0010] Based on the predicted values ​​of effluent water quality and energy consumption, controlled disturbance calculations are performed on the corresponding energy consumption, the water quality improvement response per unit energy consumption is calculated, and the energy consumption-water quality inverse contribution decomposition is performed to obtain the energy consumption-water quality contribution coefficient sequence and construct the equivalent minimum necessary energy consumption structure.

[0011] Based on the current operating parameters, hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters, a water quality reaction reachable domain is constructed. The consistency between the predicted effluent water quality and the water quality reaction reachable domain is determined, and a feasible control range that meets the mechanism constraints is identified.

[0012] Within the constraints of the equivalent minimum necessary energy consumption structure and the feasible control range, the optimal combination of control parameters is obtained by synergistic optimization of aeration rate, reflux ratio and reagent dosage. The optimal combination of control parameters is then subjected to energy consumption response inertia constraint processing.

[0013] Steady-state boundary verification is performed on the optimal control parameter combination after energy consumption response inertial constraint treatment. When the constraints of dissolved oxygen range, sludge concentration stability and effluent water quality compliance are met, the optimal control parameter combination is input into the wastewater treatment automatic control system for execution, and the multi-dimensional operating state feature matrix is ​​updated based on the execution feedback data.

[0014] Optionally, the multi-source operating data includes influent water quality parameters, influent flow rate parameters, control operating parameters, process status parameters, sub-item energy consumption parameters, and effluent water quality indicators.

[0015] Optionally, the preprocessing of multi-source operational data includes outlier removal, missing data completion, normalization, and time-series reconstruction based on a sliding time window.

[0016] Optionally, the output includes the predicted effluent water quality and the predicted energy consumption for each component within the prediction period, including:

[0017] Continuous time segments are extracted from the multidimensional operating state feature matrix according to the preset input window length as input samples. The input samples include influent water quality parameter sequence, influent flow rate parameter sequence, control variable sequence, process state parameter sequence and sub-item energy consumption parameter sequence. The parameter sequences of the input samples are channelized and spliced ​​according to parameter type to form a unified input tensor.

[0018] An improved MICN model is constructed, which consists of a multi-scale local context extraction layer, a global context interaction layer, and an energy consumption-water quality coupling consistency output layer, wherein:

[0019] The multi-scale local context extraction layer uses multiple sets of parallel convolutional channels with different receptive fields to extract local temporal features from a unified input tensor and output multi-scale local features.

[0020] The global context interaction layer calculates the changes in aeration rate, reflux ratio, and reagent dosage at adjacent times within the input window. It marks the times when the changes exceed a preset threshold as control events and generates an event indication sequence. The event indication sequence is used to perform gating weighting on the global context interaction process and outputs global context features.

[0021] The energy consumption-water quality coupling consistency output layer fuses multi-scale local features with global context features, generates effluent water quality prediction values ​​and individual energy consumption prediction values ​​through a unified mapping relationship, performs consistency constraint calculations on individual energy consumption prediction values ​​to obtain comprehensive energy consumption prediction values, and outputs effluent water quality prediction values, individual energy consumption prediction values, and comprehensive energy consumption prediction values ​​synchronously.

[0022] The improved MICN model is trained based on historical operating data. A joint training strategy is adopted to minimize the error in effluent quality prediction, the error in individual energy consumption prediction, and the consistency error between the comprehensive energy consumption prediction and the sum of individual energy consumption predictions.

[0023] The input sample corresponding to the current moment is input into the improved MICN model after training. Multi-scale local context extraction, event-guided global context interaction, and energy consumption-water quality coupling consistency output are performed to obtain the predicted value of effluent water quality, the predicted value of each component of energy consumption, and the predicted value of comprehensive energy consumption within the prediction period.

[0024] Optionally, obtaining the energy consumption-water quality contribution coefficient sequence and constructing the equivalent minimum necessary energy consumption structure includes:

[0025] Obtain the predicted values ​​of effluent water quality and energy consumption of each component within the prediction period under the current operating status, and save the predicted values ​​of effluent water quality and energy consumption of each component as the baseline prediction results.

[0026] Controlled disturbance schemes with preset amplitudes are set for aeration volume, reflux ratio and chemical dosage. Under each disturbance scheme, only the control variable corresponding to the target sub-item energy consumption is changed, while other control variables remain unchanged. The multi-dimensional operating state feature matrix after disturbance is input into the improved MICN model to obtain the corresponding disturbance effluent water quality prediction results and disturbance sub-item energy consumption prediction results.

[0027] The predicted effluent water quality under each disturbance scheme is compared with the baseline predicted effluent water quality to obtain the predicted change in water quality. The predicted energy consumption of each disturbance item is compared with the predicted energy consumption of the baseline item to obtain the predicted change in energy consumption of each item. Based on the water quality improvement related to the treatment target and the corresponding predicted change in energy consumption of each item, the unit energy consumption water quality improvement response of each item is determined.

[0028] Based on the water quality improvement response of each energy consumption item under each disturbance scheme and operating time, the energy consumption-water quality inverse contribution decomposition is performed. The water quality improvement within the prediction period is accumulated and normalized to obtain the energy consumption-water quality contribution coefficient of each energy consumption item. The energy consumption-water quality contribution coefficient sequence is constructed according to the contribution coefficient. The energy consumption items with contribution coefficients not lower than the preset threshold are marked as necessary energy consumption, and the energy consumption items with contribution coefficients lower than the preset threshold are marked as adjustable energy consumption.

[0029] Under the condition that the predicted effluent water quality meets the preset compliance requirements, based on the necessary energy consumption, the adjustable energy consumption is gradually reduced to the preset lower limit according to the energy consumption-water quality contribution coefficient sequence. After each adjustment, the predicted effluent water quality is verified. When further reducing any adjustable energy consumption will result in failure to meet the preset compliance requirements, the current combination of predicted sub-items of energy consumption is determined as the equivalent minimum necessary energy consumption structure.

[0030] Optionally, the construction of the water quality response reachability domain involves determining the consistency between the predicted effluent water quality value and the water quality response reachability domain, and identifying a feasible controllable range that satisfies the mechanistic constraints, including:

[0031] Obtain operating status parameters, hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters within the current control cycle;

[0032] A water quality reactivity reachability domain is constructed, which consists of a rate-of-change boundary layer and a steady-state boundary layer, wherein:

[0033] The rate of change boundary layer limits the maximum allowable decrease and increase of each effluent water quality index at adjacent times within the prediction period, while the steady-state boundary layer limits the minimum and maximum boundary values ​​that each effluent water quality index can reach under the current hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters.

[0034] The rate of change boundary layer is determined based on the statistical results of the actual rate of change of each effluent water quality index in the historical operation data of the most recent preset length. The steady-state boundary layer is determined based on the upper and lower limits of the biochemical reaction capacity corresponding to the current hydraulic retention time, sludge age, dissolved oxygen concentration and biochemical reaction rate parameters. The water quality reaction reachable domain is obtained by performing an intersection operation on the rate of change boundary layer and the steady-state boundary layer.

[0035] Obtain the predicted effluent water quality value within the prediction period, map the predicted effluent water quality value to the water quality response reachable domain at each time step, and determine whether the predicted effluent water quality value at each time step simultaneously satisfies the constraint conditions of the rate of change boundary layer and the steady-state boundary layer.

[0036] Control parameter combinations that do not meet the water quality reaction reachability constraints are eliminated, and control parameter combinations that meet both the water quality reaction reachability constraints and the preset effluent water quality compliance requirements are retained as feasible control ranges.

[0037] Optionally, the energy consumption response inertial constraint processing of the optimal control parameter combination includes:

[0038] Using the equivalent minimum necessary energy consumption structure and feasible control range as joint constraints, aeration rate, reflux ratio and reagent dosage are selected as collaborative optimization control variables, and corresponding control parameter combinations are constructed.

[0039] The multidimensional operating state feature matrix corresponding to different combinations of control parameters is input into the improved MICN model to obtain the corresponding predicted values ​​of effluent water quality and sub-item energy consumption. The combination of control parameters is iteratively screened with the minimum comprehensive energy consumption and the predicted value of effluent water quality meeting the standard requirements as the optimization criterion.

[0040] Under the conditions of satisfying the constraints of feasible control range and equivalent minimum necessary energy consumption structure, the aeration volume, reflux ratio and chemical dosage are adjusted and searched simultaneously to gradually narrow the search range of control parameter combination. When further adjusting any control variable will cause the predicted effluent water quality to fail to meet the standard or the comprehensive energy consumption will no longer decrease, the current control parameter combination is determined as the optimal control parameter combination.

[0041] The pool volume, hydraulic retention time, equipment response time, and biochemical reaction lag time of the wastewater treatment device are obtained. Based on the parameters, the maximum allowable variation of the control variables within adjacent control cycles is determined, and the maximum variation is used as the energy consumption response inertial constraint boundary.

[0042] The optimal control parameter combination is compared with the actual control parameters executed in the previous control cycle. When the difference between the two exceeds the maximum change range, the optimal control parameter combination is subjected to amplitude limiting processing, and the control parameter combination that satisfies the energy consumption response inertia constraint is output.

[0043] Optionally, the step of inputting the optimal control parameter combination into the wastewater treatment automatic control system for execution, and updating the multi-dimensional operating state feature matrix based on the execution feedback data, includes:

[0044] The control parameter combination after energy consumption response inertial constraint processing is obtained, and the control parameter combination is fused with the influent water quality parameters, influent flow parameters and process state parameters at the current time to form the corresponding execution state data;

[0045] The execution status data is input into the improved MICN model to obtain the corresponding predicted values ​​of the effluent water quality and the predicted values ​​of the energy consumption of the execution sub-items. Steady-state boundary verification is then performed based on the predicted values ​​of the effluent water quality.

[0046] When the predicted effluent water quality meets the effluent water quality standard range, the allowable dissolved oxygen range, and the stable mixed liquor sludge concentration range, the control parameter combination is input into the wastewater treatment automatic control device for execution, and the actual effluent water quality data, process status data, and sub-item energy consumption data after execution are collected as execution feedback data.

[0047] The execution feedback data is spliced ​​and updated with the newly collected multi-source operation data to reconstruct the multi-dimensional operation status feature matrix. Based on the updated multi-dimensional operation status feature matrix, rolling prediction, reverse contribution decomposition, reachability domain determination and collaborative optimization are performed in sequence.

[0048] According to an embodiment of the present invention, a wastewater treatment operation energy consumption optimization system based on deep learning includes:

[0049] The data acquisition and preprocessing module is used to collect multi-source operational data and preprocess it to construct a multi-dimensional operational status feature matrix.

[0050] The multi-scale time series prediction module is used to input the multi-dimensional operating status feature matrix into the improved MICN model and output the predicted values ​​of effluent water quality and sub-item energy consumption within the prediction period.

[0051] The energy consumption-water quality reverse contribution decomposition module is used to perform controlled disturbance calculations based on the predicted values ​​of effluent water quality and the predicted values ​​of individual energy consumption items, and to construct an equivalent minimum necessary energy consumption structure.

[0052] The water quality reaction reachability determination module is used to construct the water quality reaction reachability and determine the feasible control range corresponding to the predicted effluent water quality value.

[0053] The collaborative optimization and inertial constraint module is used to perform collaborative optimization to obtain the optimal combination of control parameters under the constraints of equivalent minimum necessary energy consumption structure and feasible control range, and to perform energy consumption response inertial constraint processing.

[0054] The steady-state verification execution module is used to perform steady-state verification on the optimal control parameters and execute control when the constraints are met, while updating the multi-dimensional operating state feature matrix.

[0055] The beneficial effects of this invention are:

[0056] This invention introduces an improved MICN model to perform multi-scale time-series modeling of multi-source operational data. Based on the predicted effluent water quality and individual energy consumption components, it performs an energy consumption-water quality inverse contribution decomposition, achieving a refined identification of the actual energy consumption effects of key control variables such as aeration, recirculation, and chemical dosing. Compared to existing optimization methods that rely solely on overall energy consumption or empirical adjustments, this invention can characterize the true contribution relationship of each individual energy consumption component to water quality compliance from the perspective of water quality improvement response per unit energy consumption. This allows for the construction of an equivalent minimum necessary energy consumption structure while ensuring stable effluent compliance, effectively reducing redundant energy input and improving the targeting and accuracy of energy consumption optimization.

[0057] This invention constructs a water quality reaction reachability domain based on parameters such as hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate. It then performs a mechanistic consistency assessment on the predicted effluent quality, ensuring that the optimization results always remain within the range allowed by actual biochemical reaction capacity. Compared to existing technologies that rely solely on data prediction while ignoring process mechanism constraints, this invention avoids generating control parameter combinations that deviate from actual reaction capacity, improves the feasibility and engineering applicability of the optimization results, and enhances operational reliability under conditions of fluctuating influent water quality and varying loads.

[0058] This invention, after obtaining the optimal combination of control parameters, further introduces energy consumption response inertial constraints and steady-state boundary verification mechanisms to limit the range of control parameter changes and perform stability assessments before execution. Then, it combines execution feedback data to continuously update the multi-dimensional operating state feature matrix, forming a closed-loop energy consumption optimization control process. This effectively suppresses system oscillations caused by sudden changes in control parameters, reduces dissolved oxygen fluctuations and the risk of sludge instability, and achieves a synergistic improvement in energy-saving operation and stable compliance under complex non-stationary operating conditions, thereby comprehensively improving the stability, safety, and energy efficiency of wastewater treatment operations. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is a flowchart of a wastewater treatment operation energy consumption optimization method based on deep learning proposed in this invention;

[0061] Figure 2 This is a schematic diagram of the structure of a wastewater treatment operation energy consumption optimization system based on deep learning proposed in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1 A deep learning-based method for optimizing energy consumption in wastewater treatment operations includes:

[0064] Collect multi-source operational data from the wastewater treatment system, preprocess the multi-source operational data, and form a multi-dimensional operational status feature matrix;

[0065] The multi-dimensional operating state feature matrix is ​​input into the improved MICN model to perform multi-scale time series modeling of control variables, sub-item energy consumption and process state parameters, and output the predicted values ​​of effluent water quality and sub-item energy consumption in the corresponding prediction period.

[0066] Based on the predicted values ​​of effluent water quality and energy consumption, controlled disturbance calculations are performed on the corresponding energy consumption, the water quality improvement response per unit energy consumption is calculated, and the energy consumption-water quality inverse contribution decomposition is performed to obtain the energy consumption-water quality contribution coefficient sequence and construct the equivalent minimum necessary energy consumption structure.

[0067] Based on the current operating parameters, hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters, a water quality reaction reachable domain is constructed. The consistency between the predicted effluent water quality and the water quality reaction reachable domain is determined, and a feasible control range that meets the mechanism constraints is identified.

[0068] Within the constraints of the equivalent minimum necessary energy consumption structure and the feasible control range, the optimal combination of control parameters is obtained by synergistic optimization of aeration rate, reflux ratio and reagent dosage. The optimal combination of control parameters is then subjected to energy consumption response inertia constraint processing.

[0069] Steady-state boundary verification is performed on the optimal control parameter combination after energy consumption response inertial constraint treatment. When the constraints of dissolved oxygen range, sludge concentration stability and effluent water quality compliance are met, the optimal control parameter combination is input into the wastewater treatment automatic control system for execution, and the multi-dimensional operating state feature matrix is ​​updated based on the execution feedback data.

[0070] In this embodiment, the multi-source operating data includes influent water quality parameters, influent flow rate parameters, control operating parameters, process status parameters, sub-item energy consumption parameters, and effluent water quality indicators. Influent water quality parameters include chemical oxygen demand (COD), ammonia nitrogen, total nitrogen (TNI), total phosphorus (TP), and suspended solids concentration. Influent flow rate parameters include instantaneous influent flow rate, average influent flow rate, and influent flow rate change rate. Control operating parameters include aeration rate, return flow ratio, excess sludge discharge, and chemical dosage. Process status parameters include dissolved oxygen concentration, mixed liquor suspended solids concentration, sludge age, and hydraulic retention time. Sub-item energy consumption parameters include aeration system energy consumption, return pump energy consumption, booster pump energy consumption, and chemical dosing system energy consumption. Effluent water quality indicators include effluent COD, effluent ammonia nitrogen, effluent TNI, and effluent TNI.

[0071] In this embodiment, the preprocessing of multi-source operational data includes outlier removal, missing data completion, normalization, and time-series reconstruction based on a sliding time window.

[0072] In this embodiment, the output of the predicted effluent water quality and the predicted sub-item energy consumption within the corresponding prediction period includes:

[0073] Continuous time segments are extracted from the multidimensional operating state feature matrix according to the preset input window length as input samples. The input samples include influent water quality parameter sequence, influent flow rate parameter sequence, control variable sequence, process state parameter sequence and sub-item energy consumption parameter sequence. The parameter sequences of the input samples are channelized and spliced ​​according to parameter type to form a unified input tensor.

[0074] An improved MICN model is constructed, which consists of a multi-scale local context extraction layer, a global context interaction layer, and an energy consumption-water quality coupling consistency output layer, wherein:

[0075] The multi-scale local context extraction layer uses multiple sets of parallel convolutional channels with different receptive fields to extract local temporal features from a unified input tensor and output multi-scale local features.

[0076] The global context interaction layer calculates the changes in aeration rate, reflux ratio, and reagent dosage at adjacent times within the input window. It marks the times when the changes exceed a preset threshold as control events and generates an event indication sequence. The event indication sequence is used to perform gating weighting on the global context interaction process and outputs global context features.

[0077] The energy consumption-water quality coupling consistency output layer fuses multi-scale local features with global contextual features, and generates effluent water quality predictions and individual energy consumption predictions through a unified mapping relationship. Consistency constraints are applied to each individual energy consumption prediction to obtain the comprehensive energy consumption prediction. The effluent water quality prediction, individual energy consumption predictions, and comprehensive energy consumption prediction are output synchronously. Specifically, the generation of effluent water quality predictions and individual energy consumption predictions is as follows:

[0078] The fused multi-scale local features and global context features are weighted and fused according to the time dimension and feature channel dimension to form a unified coupled feature representation. The coupled feature representation is then reconstructed based on control variables, process state parameters and historical energy consumption trends.

[0079] The coupled feature representation is input into the dual-branch mapping structure, and water quality mapping relationship and energy consumption mapping relationship are established in the same feature space. The water quality mapping relationship is used to extract feature components related to pollutant removal to generate effluent water quality prediction value, and the energy consumption mapping relationship is used to extract feature components related to aeration, reflux and chemical dosing control to generate energy consumption prediction value for each item.

[0080] Based on the coupling relationship between the predicted values ​​of each sub-item energy consumption and the control variables, the predicted values ​​of each sub-item energy consumption are corrected for structural consistency, and the corrected predicted values ​​of effluent water quality, predicted values ​​of each sub-item energy consumption, and the comprehensive predicted value of energy consumption obtained by summing up the predicted values ​​of each sub-item energy consumption are output simultaneously.

[0081] The improved MICN model is trained based on historical operating data. A joint training strategy is adopted to minimize the error in effluent quality prediction, the error in individual energy consumption prediction, and the consistency error between the comprehensive energy consumption prediction and the sum of individual energy consumption predictions.

[0082] The input sample corresponding to the current moment is input into the improved MICN model after training. Multi-scale local context extraction, event-guided global context interaction, and energy consumption-water quality coupling consistency output are performed to obtain the predicted value of effluent water quality, the predicted value of each component of energy consumption, and the predicted value of comprehensive energy consumption within the prediction period.

[0083] In this embodiment, obtaining the energy consumption-water quality contribution coefficient sequence and constructing an equivalent minimum necessary energy consumption structure includes:

[0084] Obtain the predicted values ​​of effluent water quality and energy consumption of each component within the prediction period under the current operating status, and save the predicted values ​​of effluent water quality and energy consumption of each component as the baseline prediction results.

[0085] Controlled disturbance schemes with preset amplitudes were set for aeration rate, reflux ratio, and reagent dosage. Under each disturbance scheme, only the control variable corresponding to the target energy consumption component was changed, while other control variables remained unchanged. The multidimensional operating state feature matrix after disturbance was input into the improved MICN model to obtain the corresponding effluent water quality prediction results and disturbance component energy consumption prediction results. The controlled disturbance scheme with preset amplitudes is as follows:

[0086] Aeration volume disturbance range: Based on the current control cycle aeration volume, apply a relative disturbance of ±3%, and the single disturbance shall not exceed the upper limit of ±5%.

[0087] Return current ratio disturbance amplitude: Based on the current control cycle return current ratio, apply a relative disturbance of ±2%, and the single disturbance shall not exceed the upper limit of ±4%.

[0088] The perturbation range of the dosage of the pesticide is: based on the dosage of the pesticide in the current control cycle, a relative perturbation of ±1% is applied, and the single perturbation shall not exceed the upper limit of ±3%.

[0089] The predicted effluent water quality under each disturbance scheme is compared with the baseline predicted effluent water quality to obtain the predicted change in water quality. The predicted energy consumption of each disturbance item is compared with the predicted energy consumption of the baseline item to obtain the predicted change in energy consumption of each item. Based on the water quality improvement related to the treatment target and the corresponding predicted change in energy consumption of each item, the unit energy consumption water quality improvement response of each item is determined.

[0090] Based on the water quality improvement response of each energy consumption item under each disturbance scheme and operating time, the energy consumption-water quality inverse contribution decomposition is performed. The water quality improvement within the prediction period is accumulated and normalized to obtain the energy consumption-water quality contribution coefficient of each energy consumption item. The energy consumption-water quality contribution coefficient sequence is constructed according to the contribution coefficient. The energy consumption items with contribution coefficients not lower than the preset threshold are marked as necessary energy consumption, and the energy consumption items with contribution coefficients lower than the preset threshold are marked as adjustable energy consumption.

[0091] Under the condition that the predicted effluent water quality meets the preset compliance requirements, based on the necessary energy consumption, the adjustable energy consumption is gradually reduced to the preset lower limit according to the energy consumption-water quality contribution coefficient sequence. After each adjustment, the predicted effluent water quality is verified. When further reducing any adjustable energy consumption will result in failure to meet the preset compliance requirements, the current combination of predicted sub-items of energy consumption is determined as the equivalent minimum necessary energy consumption structure.

[0092] In this embodiment, the construction of the water quality response reachability domain involves determining the consistency between the predicted effluent water quality value and the water quality response reachability domain, and identifying a feasible controllable range that satisfies the mechanism constraints. This includes:

[0093] Obtain operating status parameters, hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters within the current control cycle;

[0094] A water quality reactivity reachability domain is constructed, which consists of a rate-of-change boundary layer and a steady-state boundary layer, wherein:

[0095] The rate of change boundary layer limits the maximum allowable decrease and increase of each effluent water quality index at adjacent times within the prediction period, while the steady-state boundary layer limits the minimum and maximum boundary values ​​that each effluent water quality index can reach under the current hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters.

[0096] The rate of change boundary layer is determined based on the statistical results of the actual rate of change of each effluent water quality index in the historical operation data of the most recent preset length. The steady-state boundary layer is determined based on the upper and lower limits of the biochemical reaction capacity corresponding to the current hydraulic retention time, sludge age, dissolved oxygen concentration and biochemical reaction rate parameters. The water quality reaction reachable domain is obtained by performing an intersection operation on the rate of change boundary layer and the steady-state boundary layer.

[0097] Obtain the predicted effluent water quality value within the prediction period, map the predicted effluent water quality value to the water quality response reachable domain at each time step, and determine whether the predicted effluent water quality value at each time step simultaneously satisfies the constraint conditions of the rate of change boundary layer and the steady-state boundary layer.

[0098] Control parameter combinations that do not meet the water quality reaction reachability constraints are eliminated, and control parameter combinations that meet both the water quality reaction reachability constraints and the preset effluent water quality compliance requirements are retained as feasible control ranges.

[0099] In this embodiment, the energy consumption response inertial constraint processing of the optimal control parameter combination includes:

[0100] Using the equivalent minimum necessary energy consumption structure and feasible control range as joint constraints, aeration rate, reflux ratio and reagent dosage are selected as collaborative optimization control variables, and corresponding control parameter combinations are constructed.

[0101] The multidimensional operating state feature matrix corresponding to different combinations of control parameters is input into the improved MICN model to obtain the corresponding predicted values ​​of effluent water quality and sub-item energy consumption. The combination of control parameters is iteratively screened with the minimum comprehensive energy consumption and the predicted value of effluent water quality meeting the standard requirements as the optimization criterion.

[0102] Under the conditions of satisfying the constraints of feasible control range and equivalent minimum necessary energy consumption structure, the aeration volume, reflux ratio and chemical dosage are adjusted and searched simultaneously to gradually narrow the search range of control parameter combination. When further adjusting any control variable will cause the predicted effluent water quality to fail to meet the standard or the comprehensive energy consumption will no longer decrease, the current control parameter combination is determined as the optimal control parameter combination.

[0103] The pool volume, hydraulic retention time, equipment response time, and biochemical reaction lag time of the wastewater treatment device are obtained. Based on the parameters, the maximum allowable variation of the control variables within adjacent control cycles is determined, and the maximum variation is used as the energy consumption response inertial constraint boundary.

[0104] The optimal control parameter combination is compared with the actual control parameters executed in the previous control cycle. When the difference between the two exceeds the maximum change range, the optimal control parameter combination is subjected to amplitude limiting processing, and the control parameter combination that satisfies the energy consumption response inertia constraint is output.

[0105] In this embodiment, the step of inputting the optimal control parameter combination into the wastewater treatment automatic control system for execution, and updating the multi-dimensional operating state feature matrix based on the execution feedback data, includes:

[0106] The control parameter combination after energy consumption response inertial constraint processing is obtained, and the control parameter combination is fused with the influent water quality parameters, influent flow parameters and process state parameters at the current time to form the corresponding execution state data;

[0107] The execution status data is input into the improved MICN model to obtain the corresponding predicted values ​​of the effluent water quality and the predicted values ​​of the energy consumption of the execution sub-items. Steady-state boundary verification is then performed based on the predicted values ​​of the effluent water quality.

[0108] When the predicted effluent water quality meets the effluent water quality standard range, the allowable dissolved oxygen range, and the stable mixed liquor sludge concentration range, the control parameter combination is input into the wastewater treatment automatic control device for execution, and the actual effluent water quality data, process status data, and sub-item energy consumption data after execution are collected as execution feedback data.

[0109] The execution feedback data is spliced ​​and updated with the newly collected multi-source operation data to reconstruct the multi-dimensional operation status feature matrix. Based on the updated multi-dimensional operation status feature matrix, rolling prediction, reverse contribution decomposition, reachability domain determination and collaborative optimization are performed in sequence.

[0110] refer to Figure 2 A wastewater treatment operation energy consumption optimization system based on deep learning includes the following modules:

[0111] The data acquisition and preprocessing module is used to collect multi-source operational data and preprocess it to construct a multi-dimensional operational status feature matrix.

[0112] The multi-scale time series prediction module is used to input the multi-dimensional operating status feature matrix into the improved MICN model and output the predicted values ​​of effluent water quality and sub-item energy consumption within the prediction period.

[0113] The energy consumption-water quality reverse contribution decomposition module is used to perform controlled disturbance calculations based on the predicted values ​​of effluent water quality and the predicted values ​​of individual energy consumption items, and to construct an equivalent minimum necessary energy consumption structure.

[0114] The water quality reaction reachability determination module is used to construct the water quality reaction reachability and determine the feasible control range corresponding to the predicted effluent water quality value.

[0115] The collaborative optimization and inertial constraint module is used to perform collaborative optimization to obtain the optimal combination of control parameters under the constraints of equivalent minimum necessary energy consumption structure and feasible control range, and to perform energy consumption response inertial constraint processing.

[0116] The steady-state verification execution module is used to perform steady-state verification on the optimal control parameters and execute control when the constraints are met, while updating the multi-dimensional operating state feature matrix.

[0117] Example 1:

[0118] To verify the feasibility of this invention in practice, it was applied to a municipal wastewater treatment plant with a designed treatment capacity of 100,000 m³ / d. The plant employs an A² / O biological treatment process and serves a mixed influent of domestic sewage and a small amount of pre-treated industrial wastewater. During long-term operation, the plant exhibits typical daily fluctuations in influent quality and flow rate. On weekdays, the influent flow rate is approximately 92,000–105,000 m³ / d during the day and decreases to 76,000–84,000 m³ / d at night. The influent chemical oxygen demand (COD) fluctuates between 220 and 360 mg / L, and ammonia nitrogen fluctuates between 22 and 38 mg / L. The original operating method relied primarily on manual experience to set aeration rates and return ratios. To ensure stable effluent compliance, a high aeration intensity was typically maintained, even during low-load periods, resulting in high aeration energy consumption. Due to frequent manual adjustments, dissolved oxygen fluctuated significantly, and the mixed liquor sludge concentration was unstable. Furthermore, it was difficult to identify the actual effective energy consumption contribution of each control unit, redundant energy input, and significant control fluctuations.

[0119] The deep learning-based energy consumption optimization method for wastewater treatment operation described in this invention is deployed in this wastewater treatment plant. First, data from the plant's online monitoring and SCADA systems is accessed via a data acquisition and preprocessing module. This involves continuously collecting data on influent water quality parameters, influent flow rate parameters, aeration volume, return ratio, reagent dosage, dissolved oxygen, mixed liquor sludge concentration, sludge age, hydraulic retention time, and the energy consumption data of aeration blowers and electric return pumps. Outlier removal, missing value completion, and normalization are performed on the multi-source operational data to form a multi-dimensional operational state feature matrix. Subsequently, this multi-dimensional operational state feature matrix is ​​input into an improved MICN model to perform multi-scale time-series modeling of control variables, energy consumption components, and process state parameters, outputting predicted effluent water quality and predicted energy consumption values ​​for the next two hours. Controlled disturbance calculations are performed on the corresponding sub-items of energy consumption. By comparing the predicted changes in effluent water quality before and after the disturbance with the changes in sub-items of energy consumption, the water quality improvement response per unit of energy consumption is calculated, and the energy consumption-water quality inverse contribution decomposition is performed to obtain the energy consumption-water quality contribution coefficient sequence. Based on this, an equivalent minimum necessary energy consumption structure is constructed to make the input of aeration energy consumption and return flow energy consumption more in line with the actual water quality improvement needs.

[0120] During the optimization process, a water quality reaction reachable domain was constructed based on current operating parameters, a hydraulic retention time of approximately 7.2 h, a sludge age of approximately 13.8 d, a dissolved oxygen control range of 1.8–2.5 mg / L, and historical biochemical reaction rate data. The effluent water quality predictions output by the improved MICN model were then compared with this water quality reaction reachable domain for consistency assessment, retaining only feasible control intervals that satisfy the mechanistic constraints. Within the equivalent minimum necessary energy consumption structural constraints and feasible control intervals, the aeration rate, recirculation ratio, and reagent dosage were synergistically optimized to obtain the optimal control parameter combination. Energy consumption response inertia constraints were then applied to this optimal control parameter combination based on tank volume, hydraulic retention time, and equipment response time to limit the variation in aeration rate and recirculation ratio within adjacent control cycles.

[0121] Subsequently, the optimal control parameter combination after inertial constraint processing is subjected to steady-state boundary verification. When the constraints of dissolved oxygen range, sludge concentration stability and effluent water quality compliance are met, the optimal control parameter combination is input into the wastewater treatment automatic control system for execution, and the multi-dimensional operating state feature matrix is ​​updated based on the execution feedback data to achieve rolling prediction and closed-loop optimization control.

[0122] In a 45-day comparative analysis of actual operation, typical operating data were selected for statistical analysis. Under the condition that the influent water quality fluctuations were basically consistent, after applying the method of this invention, the aeration rate was dynamically adjusted according to the load changes, avoiding excessive aeration during low-load periods. Through the energy consumption-water quality reverse contribution decomposition, it was identified that the actual effect of recirculation energy consumption on ammonia nitrogen removal was relatively stable, allowing the optimization process to prioritize adjusting the aeration intensity rather than blindly increasing the recirculation ratio.

[0123] Table 1. Comparison of key operating data before and after energy consumption optimization under actual operating conditions.

[0124]

[0125] As can be seen from the data in Table 1, the average influent COD in the first and second weeks before optimization was 312 mg / L and 285 mg / L, respectively, with influent flow rates of 95800 m³ / d and 92100 m³ / d. The corresponding average aeration power consumption was 7420 kWh / d and 7010 kWh / d, respectively, and the total operating power consumption was 12060 kWh / d and 11640 kWh / d, respectively. The effluent ammonia nitrogen was 3.9 mg / L and 4.1 mg / L, respectively. This indicates that under the traditional operation mode, a high aeration intensity is usually maintained to ensure compliance with standards, and energy consumption passively changes with load fluctuations, and the stability of effluent indicators is generally poor.

[0126] After applying the method of this invention, under similar operating conditions with influent COD of 276–305 mg / L and influent flow rate of 90,800–97,200 m³ / d, the average aeration power consumption decreased to 6,320–6,710 kWh / d, and the total operating power consumption decreased to 10,320–10,920 kWh / d, showing a continuous downward trend. This indicates that by improving MICN prediction and constructing an equivalent minimum necessary energy consumption structure through energy consumption-water quality inverse contribution decomposition, redundant energy consumption can be effectively reduced without changing the treatment load level.

[0127] After optimization, the ammonia nitrogen in the effluent remained stable at 2.7–3.2 mg / L each week, which was significantly lower than the previous range of 3.9–4.1 mg / L and the fluctuation was reduced. This indicates that the synergistic optimization under the constraints of the feasible control range and the water quality response reachable domain, combined with the energy consumption response inertia constraint treatment, not only reduced the operating energy consumption but also improved the stability of the effluent. This verifies the actual effect of the present invention in achieving stable compliance and energy-saving operation under fluctuating influent conditions.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing energy consumption in wastewater treatment operations based on deep learning, characterized in that, include: Collect multi-source operational data from the wastewater treatment system, preprocess the multi-source operational data, and form a multi-dimensional operational status feature matrix; The multi-dimensional operating state feature matrix is ​​input into the improved MICN model to perform multi-scale time series modeling of control variables, sub-item energy consumption and process state parameters, and output the predicted values ​​of effluent water quality and sub-item energy consumption in the corresponding prediction period. Based on the predicted values ​​of effluent water quality and energy consumption, controlled disturbance calculations are performed on the corresponding energy consumption, the water quality improvement response per unit energy consumption is calculated, and the energy consumption-water quality inverse contribution decomposition is performed to obtain the energy consumption-water quality contribution coefficient sequence and construct the equivalent minimum necessary energy consumption structure. Based on the current operating parameters, hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters, a water quality reaction reachable domain is constructed. The consistency between the predicted effluent water quality and the water quality reaction reachable domain is determined, and a feasible control range that meets the mechanism constraints is identified. Within the constraints of the equivalent minimum necessary energy consumption structure and the feasible control range, the optimal combination of control parameters is obtained by synergistic optimization of aeration rate, reflux ratio and reagent dosage. The optimal combination of control parameters is then subjected to energy consumption response inertia constraint processing. Steady-state boundary verification is performed on the optimal control parameter combination after energy consumption response inertial constraint treatment. When the constraints of dissolved oxygen range, sludge concentration stability and effluent water quality compliance are met, the optimal control parameter combination is input into the wastewater treatment automatic control system for execution, and the multi-dimensional operating state feature matrix is ​​updated based on the execution feedback data.

2. The wastewater treatment operation energy consumption optimization method based on deep learning according to claim 1, characterized in that, The multi-source operation data includes influent water quality parameters, influent flow rate parameters, control operation parameters, process status parameters, sub-item energy consumption parameters, and effluent water quality indicators.

3. The method for optimizing wastewater treatment operation energy consumption based on deep learning according to claim 1, characterized in that, The preprocessing of multi-source operational data includes outlier removal, missing data completion, normalization, and time series reconstruction based on a sliding time window.

4. The wastewater treatment operation energy consumption optimization method based on deep learning according to claim 1, characterized in that, The output corresponds to the predicted values ​​of effluent water quality and energy consumption for the prediction period, including: Continuous time segments are extracted from the multidimensional operating state feature matrix according to the preset input window length as input samples. The input samples include influent water quality parameter sequence, influent flow rate parameter sequence, control variable sequence, process state parameter sequence and sub-item energy consumption parameter sequence. The parameter sequences of the input samples are channelized and spliced ​​according to parameter type to form a unified input tensor. An improved MICN model is constructed, which consists of a multi-scale local context extraction layer, a global context interaction layer, and an energy consumption-water quality coupling consistency output layer, wherein: The multi-scale local context extraction layer uses multiple sets of parallel convolutional channels with different receptive fields to extract local temporal features from a unified input tensor and output multi-scale local features. The global context interaction layer calculates the changes in aeration rate, reflux ratio, and reagent dosage at adjacent times within the input window. It marks the times when the changes exceed a preset threshold as control events and generates an event indication sequence. The event indication sequence is used to perform gating weighting on the global context interaction process and outputs global context features. The energy consumption-water quality coupling consistency output layer fuses multi-scale local features with global context features, generates effluent water quality prediction values ​​and individual energy consumption prediction values ​​through a unified mapping relationship, performs consistency constraint calculations on individual energy consumption prediction values ​​to obtain comprehensive energy consumption prediction values, and outputs effluent water quality prediction values, individual energy consumption prediction values, and comprehensive energy consumption prediction values ​​synchronously. The improved MICN model is trained based on historical operating data. A joint training strategy is adopted to minimize the error in effluent quality prediction, the error in individual energy consumption prediction, and the consistency error between the comprehensive energy consumption prediction and the sum of individual energy consumption predictions. The input sample corresponding to the current moment is input into the improved MICN model after training. Multi-scale local context extraction, event-guided global context interaction, and energy consumption-water quality coupling consistency output are performed to obtain the predicted value of effluent water quality, the predicted value of each component of energy consumption, and the predicted value of comprehensive energy consumption within the prediction period.

5. The wastewater treatment operation energy consumption optimization method based on deep learning according to claim 1, characterized in that, The process of obtaining the energy consumption-water quality contribution coefficient sequence and constructing an equivalent minimum necessary energy consumption structure includes: Obtain the predicted values ​​of effluent water quality and energy consumption of each component within the prediction period under the current operating status, and save the predicted values ​​of effluent water quality and energy consumption of each component as the baseline prediction results. Controlled disturbance schemes with preset amplitudes are set for aeration volume, reflux ratio and chemical dosage. Under each disturbance scheme, only the control variable corresponding to the target sub-item energy consumption is changed, while other control variables remain unchanged. The multi-dimensional operating state feature matrix after disturbance is input into the improved MICN model to obtain the corresponding disturbance effluent water quality prediction results and disturbance sub-item energy consumption prediction results. The predicted effluent water quality under each disturbance scheme is compared with the baseline predicted effluent water quality to obtain the predicted change in water quality. The predicted energy consumption of each disturbance item is compared with the predicted energy consumption of the baseline item to obtain the predicted change in energy consumption of each item. Based on the water quality improvement related to the treatment target and the corresponding predicted change in energy consumption of each item, the unit energy consumption water quality improvement response of each item is determined. Based on the water quality improvement response of each energy consumption item under each disturbance scheme and operating time, the energy consumption-water quality inverse contribution decomposition is performed. The water quality improvement within the prediction period is accumulated and normalized to obtain the energy consumption-water quality contribution coefficient of each energy consumption item. The energy consumption-water quality contribution coefficient sequence is constructed according to the contribution coefficient. The energy consumption items with contribution coefficients not lower than the preset threshold are marked as necessary energy consumption, and the energy consumption items with contribution coefficients lower than the preset threshold are marked as adjustable energy consumption. Under the condition that the predicted effluent water quality meets the preset compliance requirements, based on the necessary energy consumption, the adjustable energy consumption is gradually reduced to the preset lower limit according to the energy consumption-water quality contribution coefficient sequence. After each adjustment, the predicted effluent water quality is verified. When further reducing any adjustable energy consumption will result in failure to meet the preset compliance requirements, the current combination of predicted sub-items of energy consumption is determined as the equivalent minimum necessary energy consumption structure.

6. The wastewater treatment operation energy consumption optimization method based on deep learning according to claim 1, characterized in that, The construction of the water quality response reachability domain involves determining the consistency between the predicted effluent water quality value and the water quality response reachability domain, and identifying a feasible controllable range that satisfies the mechanism constraints, including: Obtain operating status parameters, hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters within the current control cycle; A water quality reactivity reachability domain is constructed, which consists of a rate-of-change boundary layer and a steady-state boundary layer, wherein: The rate of change boundary layer limits the maximum allowable decrease and increase of each effluent water quality index at adjacent times within the prediction period, while the steady-state boundary layer limits the minimum and maximum boundary values ​​that each effluent water quality index can reach under the current hydraulic retention time, sludge age, dissolved oxygen concentration, and biochemical reaction rate parameters. The rate of change boundary layer is determined based on the statistical results of the actual rate of change of each effluent water quality index in the historical operation data of the most recent preset length. The steady-state boundary layer is determined based on the upper and lower limits of the biochemical reaction capacity corresponding to the current hydraulic retention time, sludge age, dissolved oxygen concentration and biochemical reaction rate parameters. The water quality reaction reachable domain is obtained by performing an intersection operation on the rate of change boundary layer and the steady-state boundary layer. Obtain the predicted effluent water quality value within the prediction period, map the predicted effluent water quality value to the water quality response reachable domain at each time step, and determine whether the predicted effluent water quality value at each time step simultaneously satisfies the constraint conditions of the rate of change boundary layer and the steady-state boundary layer. Control parameter combinations that do not meet the water quality reaction reachability constraints are eliminated, and control parameter combinations that meet both the water quality reaction reachability constraints and the preset effluent water quality compliance requirements are retained as feasible control ranges.

7. The wastewater treatment operation energy consumption optimization method based on deep learning according to claim 1, characterized in that, The energy consumption response inertial constraint processing of the optimal control parameter combination includes: Using the equivalent minimum necessary energy consumption structure and feasible control range as joint constraints, aeration rate, reflux ratio and reagent dosage are selected as collaborative optimization control variables, and corresponding control parameter combinations are constructed. The multidimensional operating state feature matrix corresponding to different combinations of control parameters is input into the improved MICN model to obtain the corresponding predicted values ​​of effluent water quality and sub-item energy consumption. The combination of control parameters is iteratively screened with the minimum comprehensive energy consumption and the predicted value of effluent water quality meeting the standard requirements as the optimization criterion. Under the conditions of satisfying the constraints of feasible control range and equivalent minimum necessary energy consumption structure, the aeration volume, reflux ratio and chemical dosage are adjusted and searched simultaneously to gradually narrow the search range of control parameter combination. When further adjusting any control variable will cause the predicted effluent water quality to fail to meet the standard or the comprehensive energy consumption will no longer decrease, the current control parameter combination is determined as the optimal control parameter combination. The pool volume, hydraulic retention time, equipment response time, and biochemical reaction lag time of the wastewater treatment device are obtained. Based on the parameters, the maximum allowable variation of the control variables within adjacent control cycles is determined, and the maximum variation is used as the energy consumption response inertial constraint boundary. The optimal control parameter combination is compared with the actual control parameters executed in the previous control cycle. When the difference between the two exceeds the maximum change range, the optimal control parameter combination is subjected to amplitude limiting processing, and the control parameter combination that satisfies the energy consumption response inertia constraint is output.

8. The method for optimizing wastewater treatment operation energy consumption based on deep learning according to claim 1, characterized in that, The step of inputting the optimal control parameter combination into the wastewater treatment automatic control system for execution, and updating the multi-dimensional operating state feature matrix based on the execution feedback data, includes: The control parameter combination after energy consumption response inertial constraint processing is obtained, and the control parameter combination is fused with the influent water quality parameters, influent flow parameters and process state parameters at the current time to form the corresponding execution state data; The execution status data is input into the improved MICN model to obtain the corresponding predicted values ​​of the effluent water quality and the predicted values ​​of the energy consumption of the execution sub-items. Steady-state boundary verification is then performed based on the predicted values ​​of the effluent water quality. When the predicted effluent water quality meets the effluent water quality standard range, the allowable dissolved oxygen range, and the stable mixed liquor sludge concentration range, the control parameter combination is input into the wastewater treatment automatic control device for execution, and the actual effluent water quality data, process status data, and sub-item energy consumption data after execution are collected as execution feedback data. The execution feedback data is spliced ​​and updated with the newly collected multi-source operation data to reconstruct the multi-dimensional operation status feature matrix. Based on the updated multi-dimensional operation status feature matrix, rolling prediction, reverse contribution decomposition, reachability domain determination and collaborative optimization are performed in sequence.

9. A wastewater treatment operation energy consumption optimization system based on deep learning, comprising executing the wastewater treatment operation energy consumption optimization method based on deep learning as described in any one of claims 1 to 8, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-source operational data and preprocess it to construct a multi-dimensional operational status feature matrix. The multi-scale time series prediction module is used to input the multi-dimensional operating status feature matrix into the improved MICN model and output the predicted values ​​of effluent water quality and sub-item energy consumption within the prediction period. The energy consumption-water quality reverse contribution decomposition module is used to perform controlled disturbance calculations based on the predicted values ​​of effluent water quality and the predicted values ​​of individual energy consumption items, and to construct an equivalent minimum necessary energy consumption structure. The water quality reaction reachability determination module is used to construct the water quality reaction reachability and determine the feasible control range corresponding to the predicted effluent water quality value. The collaborative optimization and inertial constraint module is used to perform collaborative optimization to obtain the optimal combination of control parameters under the constraints of equivalent minimum necessary energy consumption structure and feasible control range, and to perform energy consumption response inertial constraint processing. The steady-state verification execution module is used to perform steady-state verification on the optimal control parameters and execute control when the constraints are met, while updating the multi-dimensional operating state feature matrix.

Images