Sewage treatment dosing method and system based on machine learning, electronic equipment, storage medium and product

Abstract

The invention discloses a sewage treatment dosing method and system based on machine learning, electronic equipment, a storage medium and a product. The method comprises the following steps: constructing a multi-parameter time sequence data set between historical operation data about a sewage treatment system and historical dosing amount data of time lag duration t corresponding to each group of historical operation data; a dosing prediction model is constructed based on a random forest network, and a multi-parameter time sequence data set is used for training; dynamically updating the lagging duration t; operating parameters of the sewage treatment system are collected in real time, and the trained dosing prediction model is used for predicting the dosing amount of the optimal dosing lagging duration. According to the method, the hysteresis of an existing dosing control mode is effectively compensated, meanwhile, the delay time is dynamically regulated and controlled, accurate prediction of the dosing amount is achieved, the fluctuation risk of the effluent quality is reduced, and the working condition generalization ability of a dosing system is improved.

Description

Technical Field

[0001] This invention relates to the field of wastewater treatment technology, and in particular to a wastewater treatment dosing method, system, electronic device, storage medium, and product based on machine learning. Background Technology

[0002] The dosing of chemicals in wastewater treatment processes is a core control step for achieving efficient pollutant removal, and its control precision directly affects the effluent quality compliance rate and operating cost control. For example, in total phosphorus removal processes, the dosage of chemicals such as polyaluminum chloride (PAC) and ferric chloride (FeCl3) needs to be dynamically matched with parameters such as the total phosphorus concentration and flow rate of the influent. Insufficient dosing will lead to excessive total phosphorus in the effluent, while excessive dosing will increase the cost of chemical procurement, raise the cost of subsequent sludge treatment and disposal, and may also cause secondary problems such as increased effluent turbidity.

[0003] Currently, wastewater treatment plant chemical dosing control methods include open-loop control based on fixed parameters, where the dosing pump frequency or chemical dosage is preset based on historical experience, and manual adjustments are only made when abnormal water quality is detected during routine inspections. This method relies entirely on the experience of maintenance personnel and is slow to respond to dynamic fluctuations in influent water quality and flow rate. Closed-loop control based on simple feedback, such as using a proportional-integral-derivative (PID) algorithm, adjusts the dosing dosage based on the effluent pollutant concentration as a feedback signal. However, pollutant removal involves nonlinear reactions coupled with multiple stages, including physical sedimentation, chemical flocculation, and biodegradation. The linear adjustment logic based on PID control is still difficult to fundamentally adapt to this time-delayed nonlinear system, often resulting in "overshoot" or "regulation oscillation," which neither guarantees water quality stability nor avoids chemical waste.

[0004] Therefore, in view of the shortcomings of existing dosing systems in terms of time-delay adaptability, dynamic adjustment accuracy and operating condition generalization ability, there is an urgent need for a machine learning dosing method that can accurately match the time-delay characteristics of wastewater treatment systems and integrate the advantages of human experience and data-driven approaches, so as to achieve dynamic and precise dosing of chemicals and balance the effluent water quality to meet standards. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a wastewater treatment dosing method, system, electronic device, storage medium, and product based on machine learning, which dynamically adjusts the lag time to improve the accuracy of dosing control and the stability of effluent quality.

[0006] Technical Solution: To achieve the above objectives, the present invention provides a wastewater treatment dosing method based on machine learning, comprising the following steps:

[0007] S1. Within a historical time period, collect multiple sets of operation data of the sewage treatment system at preset time intervals, and obtain the historical dosage data of each set of operation data corresponding to the time lag time t, and construct a multi-parameter time series dataset.

[0008] S2. Construct a drug dosing prediction model based on a random forest network and train it using a multi-parameter time series dataset;

[0009] The initial t value is the physical residence time t0 of wastewater in the wastewater treatment system, and the t value is updated as follows:

[0010] Using t0 as a baseline, generate several sets of candidate values ​​within a preset time window. For each candidate value t i Historical operating data of the wastewater treatment system is obtained, input into a pre-trained dosing prediction model, and the lag time t is obtained. i +1 Predicted dosage i ; based on the corresponding predicted dosage and the predicted dosage The actual dosage at the same time With the objectives of minimizing error and maximizing the compliance rate of total phosphorus in effluent, the optimal lag time t value is selected.

[0011] S3. Real-time acquisition of sewage treatment system operating parameters, and prediction of the optimal dosing dosage with optimal dosing lag time using the trained dosing prediction model.

[0012] Preferably, the multi-parameter time-series dataset includes a time-series matrix of historical operational data from time nt to time n and a historical drug dosage time-series matrix covering time n+1 to time n+t+1, respectively represented as:

[0013] ,

[0014] Where n represents the index of the discrete time point, L represents the influent flow rate, InfCOD represents the influent COD concentration, InfAN represents the influent ammonia nitrogen, InfTN represents the influent total nitrogen concentration, InfTP represents the influent total phosphorus, MLSS represents the mixed liquor suspended solids concentration of the biological treatment tank, EffTP represents the effluent total phosphorus concentration, and S represents the dosage.

[0015] Preferably, historical operating data and historical dosage are processed differently to form time-series data with a unified time scale.

[0016] Preferably, the differentiated treatment method is as follows: the dosage S, influent flow rate L, influent ammonia nitrogen InfAN, MLSS and pH value are collected at minute intervals to form an original minute-level time series;

[0017] The influent total nitrogen concentration (InfTN), total phosphorus concentration (InfTP), and COD concentration (InfCOD) were collected at hourly intervals. The time gap between adjacent collection points was filled by linear interpolation, and the data were converted into a minute-level time series.

[0018] The total phosphorus concentration (EffTP) in the effluent is calculated by averaging the collected values ​​hourly. The hourly average of EffTP is then mapped to all minute-level time points within that hour to generate a continuous minute-level time series.

[0019] Preferably, to address the fluctuations and dynamic changes in actual water quality, the physical residence time of wastewater in the wastewater treatment system is used as the baseline value t0. Within a preset time window, several candidate values ​​are generated based on t0. For each candidate value t... i The generation method is as follows: select several discrete candidate t values ​​within the interval [t0-Δt, t0+Δt], denoted as: t i =t0±kT, k=0,1,2,…,K, T represents the acquisition time step, K represents the total number of acquisitions, kT≤Δt, i represents the i-th candidate t value, i=0,1,2,…,K, Δt represents the time offset.

[0020] Preferably, by constructing a reward function, for each candidate value t i Corresponding predicted dosage i and the predicted dosage The actual dosage S at the same moment i The error and the compliance rate of total phosphorus in the effluent are evaluated. The optimal lag time t is selected by maximizing the reward function, which is expressed as:

[0021] ,

[0022] Where P is the lag time t i The total phosphorus compliance rate of the effluent is calculated using α and β as weighting coefficients (satisfying α+β=1, with α ranging from 0.6 to 0.8, prioritizing prediction accuracy).

[0023] The wastewater treatment dosing system based on machine learning described in this invention includes the following modules:

[0024] Multi-parameter time series dataset construction module: used to collect multiple sets of sewage treatment system operation data at preset time intervals within a historical time period, and obtain the historical dosing data of the time lag t corresponding to each set of operation data to construct a multi-parameter time series dataset;

[0025] Model building module: used to build a drug dosing prediction model based on a random forest network and train it using a multi-parameter time series dataset;

[0026] The t-value update module: Using the physical residence time t0 of wastewater in the wastewater treatment system as the initial t-value, it generates several sets of candidate values ​​within a preset time window, with t0 as the baseline. For each candidate value t... i Historical operating data of the wastewater treatment system is obtained, input into a pre-trained dosing prediction model, and the lag time t is obtained. i +1 Predicted dosage i ; based on the corresponding predicted dosage and the predicted dosage The actual dosage at the same time With the objectives of minimizing error and maximizing the compliance rate of total phosphorus in effluent, the optimal lag time t value is selected.

[0027] Real-time dosing prediction module: Collects operating parameters of the wastewater treatment system in real time and uses the trained dosing prediction model to predict the dosing dosage with the optimal dosing lag time.

[0028] An electronic device according to the present invention includes: at least one processor, and a memory communicatively connected to the at least one processor;

[0029] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the above-described machine learning-based wastewater treatment dosing method.

[0030] The present invention provides a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute the above-described machine learning-based wastewater treatment dosing method.

[0031] The present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the aforementioned wastewater treatment dosing method based on machine learning.

[0032] Beneficial Effects: This invention has the following advantages: 1. Addressing the lag problem in existing dosing control methods, this invention constructs a dataset of time-displaced sewage treatment system operation data and dosing dosage based on the lag time from the influent measurement point to the dosing point. This enables the dosing prediction model to learn the lag relationship between sewage treatment system operation data and dosing dosage, and to dynamically adjust the lag time to achieve accurate prediction of dosing dosage; 2. Based on the powerful generalization ability of machine learning, this invention can extract the inherent laws of sewage treatment system operation data and dosing dosage under complex operating conditions from a large amount of operational data to achieve dynamic adjustment of dosing dosage, reducing the risk of effluent water quality fluctuations caused by response delay. Attached Figure Description

[0033] Figure 1 This is a block diagram of a wastewater treatment system.

[0034] Figure 2 This is a flowchart illustrating the method.

[0035] Figure 3 This is a trend chart of total phosphorus in the effluent under normal operating conditions of a wastewater treatment system, based on the method of this invention for adjusting the dosage. Detailed Implementation

[0036] The technical solution of the present invention will be described in detail below with reference to the embodiments and accompanying drawings.

[0037] Example 1

[0038] like Figure 1 As shown, this embodiment provides a wastewater treatment system, including a biological treatment tank 1, a biological treatment tank effluent water quality parameter detection point 2, a wastewater transmission pipeline 3, a server 4, a user terminal 5, a mixing lift pump 6, a mixing lift tank 7, a coagulation sedimentation tank influent total phosphorus analyzer 8, a dosing pump station 9, a phosphorus removal agent PAC flow meter 10, a phosphorus removal agent PAC storage tank 11, a phosphorus removal agent PAC dosing pump 12, a coagulation sedimentation tank dosing point 13, and a coagulation sedimentation tank effluent total phosphorus analyzer 14. Polyaluminum chloride (PAC, concentration 10%) is added at the coagulation sedimentation tank dosing point 13 via the phosphorus removal agent PAC dosing pump 12 with frequency conversion function, removing total phosphorus by chemical precipitation.

[0039] The wastewater quality and system operating parameters of the wastewater treatment system are as follows: the chemical oxygen demand (COD) concentration of the influent to the biological treatment tank ranges from 100 to 300 mg / L, the ammonia nitrogen concentration ranges from 20 to 40 mg / L, the total phosphorus concentration ranges from 3 to 8 mg / L, the sludge concentration (MLSS) of the biological treatment tank is 4200 to 5400 mg / L, the pH value of the aeration tank is 6.5 to 8.5, and the influent flow rate fluctuates from 1500 to 2500 m³ / h.

[0040] A water quality analyzer and an electromagnetic flowmeter are installed at point 2, the effluent water quality parameter monitoring point of the biological treatment tank. Real-time data are collected from the outlet of biological treatment tank 1 to collect the concentrations of pollutants such as COD (InfCOD), ammonia nitrogen (InfAN), and total nitrogen (InfTN) in the influent before it enters the coagulation sedimentation tank, as well as the wastewater flow rate (L). Sensors are deployed inside biological treatment tank 1 to monitor the sludge mixed liquor suspended solids concentration (MLSS) and pH value. Wastewater from biological treatment tank 1 enters the mixing and lifting tank 7 via transmission pipe 3, and is then mixed and lifted... Pump 6 lifts the sewage to coagulation sedimentation tank 13. The total phosphorus detector 8 in the coagulation sedimentation tank influent measures the total phosphorus concentration (InfTP) in the coagulation sedimentation tank. The phosphorus removal agent PAC dosing pump 12 adds phosphorus removal agent to the coagulation zone of the coagulation sedimentation tank, and the dosage (S) of the phosphorus removal agent PAC dosing pump 12 is recorded in real time by the phosphorus removal agent PAC flow meter 10. At the effluent end of the coagulation sedimentation tank, the total phosphorus concentration (EffTP) in the effluent of the sewage treatment system is measured by the total phosphorus detector 14 in the effluent of the coagulation sedimentation tank to verify the treatment effect of the dosing.

[0041] Example 2

[0042] This embodiment, based on the wastewater treatment system of Embodiment 1, provides a wastewater treatment chemical dosing method based on machine learning, including the following:

[0043] S1. Collect historical operation data and historical chemical dosage data of the sewage treatment system, and adopt differentiated processing strategies based on parameter characteristics to form time series data with a unified time scale, thereby constructing a multi-parameter time series dataset.

[0044] The method for generating time-series data with a unified time scale by adopting a differentiated processing strategy based on parameter characteristics is as follows:

[0045] Dosage (S), influent flow rate (L), influent ammonia nitrogen (InfAN), MLSS and pH value were collected and recorded directly at minute intervals to form the raw minute-level time series;

[0046] The concentrations of total nitrogen (InfTN), total phosphorus (InfTP), and COD (InfCOD) in the influent were collected at hourly intervals. The time gaps between adjacent collection points were filled by linear interpolation to convert them into complete minute-level time series.

[0047] The total phosphorus (EffTP) concentration in the effluent is calculated by averaging the recorded values ​​hourly, mapping the hourly average of EffTP to all minute-level time points within that hour, and generating a continuous minute-level time series.

[0048] Based on time series data, a multi-parameter time series dataset is constructed. The sample format of this dataset is: the input is the running data at a certain historical moment, and the output is the dosage of medicine lagging behind that historical moment t+1 hours later.

[0049] The time series matrix covering historical operational data from time nt to time n (time range 30-60 minutes) and the historical dosage time series matrix covering historical time n+1 to time n+t+1 are respectively represented as follows:

[0050] ,

[0051] Where n represents the index of the discrete time point, and t represents the drug administration delay time.

[0052] S2. Construct a drug dosing prediction model based on a random forest network and train it using a multi-parameter time series dataset.

[0053] This embodiment utilizes the characteristics of random forest network ensemble learning to capture nonlinear correlations between parameters. During training, a sliding window method is used to transform historical time-series features into structured input, with a training data time span of n ≥ 7200. The number of decision trees in the model is set to 100, and the maximum depth is limited to 15. A 5-fold cross-validation strategy is adopted, dividing the training dataset into 5 mutually exclusive and uniformly distributed subsets. The model hyperparameters are dynamically adjusted iteratively, using 4 subsets as the training set and 1 subset as the validation set, to ensure stable generalization ability under different data distributions.

[0054] Meanwhile, the root mean square error (RMSE) is used to quantify the deviation between the predicted dosage and the actual optimal dosage. The monitoring cycle is 1 hour, and the RMSE and effluent compliance rate are calculated every hour. If the RMSE is greater than 15% or the compliance rate is less than 90% in any one of the 72 consecutive monitoring cycles, the incremental training mechanism is automatically triggered. At the same time, a full retraining is carried out every quarter to re-optimize the model structure based on the accumulated historical data, covering long-term changes such as seasonal fluctuations in water quality and drift of process parameters, to ensure that the model continuously adapts to the dynamic operation requirements of the sewage treatment system.

[0055] The update method for the drug administration lag time t value is as follows:

[0056] 1. Taking the physical residence time t0 of wastewater from the effluent detection point 2 of the biological treatment tank to the chemical dosing point 13 of the coagulation sedimentation tank as the center, several discrete candidate t values ​​are selected within the interval [t0-Δt, t0+Δt], denoted as: t i =t0±kT, k=0,1,2,…,K, T represents the acquisition time step, N represents the total number of acquisitions, i=0,1,2,…,K, nT≤Δt.

[0057] 2. Obtain all candidate t i Historical running data of the value and the corresponding t value iActual dosage after +1 time period All candidate t i The historical data corresponding to each value is input into the trained dosing prediction model to obtain each candidate t. i The historical running data corresponding to t i Predicted dosage after +1 time interval i .

[0058] The historical operating data includes: influent flow rate (L), influent COD (InfCOD), influent ammonia nitrogen (InfAN), influent total nitrogen (InfTN), influent total phosphorus (InfTP), effluent total phosphorus (EffTP), pH value, and MLSS.

[0059] 3. For each candidate t i The predicted dosage corresponding to the value i and the predicted dosage The actual dosage at the same time The objective functions are to minimize the error and maximize the compliance rate of total phosphorus (EffTP) in effluent. A reward function is constructed, and the optimal lag time t is selected by maximizing R(t).

[0060] ,

[0061] in, Lag duration t i The effluent water quality compliance rate under the +1 value (range 0~1) is calculated as follows: when the lag time is set to t, the proportion of samples with EffTP≤0.5mg / L within 1 hour to the total number of samples tested is calculated (range 0~1), and the weighting coefficient satisfies α+β=1 (α is 0.6~0.8, prioritizing the prediction accuracy).

[0062] S3. Real-time collection of sewage treatment system operating parameters, input into the trained dosing prediction model, to predict the dosing dosage for the optimal dosing lag time t.

[0063] Example 3

[0064] Taking steady-state control under normal operating conditions of a wastewater treatment system as an example, this paper provides an adjustment process to achieve a target total phosphorus concentration in the effluent of ≤0.5 mg / L using the method described in Example 2.

[0065] Based on historical system data, the optimal lag time t=25 minutes was obtained. 25 minutes prior to the current moment (26 minutes before the current moment), the influent flow rate was 2000 m³ / h, ammonia nitrogen concentration was 21 mg / L, total nitrogen concentration was 25 mg / L, total phosphorus concentration was 3.5 mg / L, MLSS was 4600 mg / L, and pH was 7.2. The optimal dosage of phosphorus removal agent PAC was predicted to be 357 L / h based on the dosing prediction model. Comparing this to the actual dosage of 335 L / h, the relative error was 6.5%, exceeding the preset 5% adjustment threshold. Therefore, the frequency of the PAC dosing pump 12 was adjusted, changing the dosage from 335 L / h to 357 L / h. After adjustment, the data from the total phosphorus analyzer 14 in the coagulation sedimentation tank effluent was continuously monitored. The results showed that the total phosphorus concentration in the water remained stable at 0.4 mg / L, meeting the requirement of effluent total phosphorus concentration ≤ 0.5 mg / L. The target requirement of mg / L was met with a 100% compliance rate. Figure 3 The data trend is shown.

Claims

1. A machine learning-based dosing method for sewage treatment, characterized by, The method comprises the following steps: S1. In a historical time period, a plurality of groups of operation data of the sewage treatment system are collected at preset time intervals, and historical dosing amount data of a lag time t+1 corresponding to each group of operation data is obtained, to construct a multi-parameter time series data set; S2. A dosing prediction model is constructed based on a random forest, and the multi-parameter time series data set is used for training; wherein the physical residence time t0 of the sewage in the sewage treatment system is taken as an initial t value, and the t value updating method is: Generate several groups of candidate values within a preset time window with t0 as the benchmark, for each candidate value t i , obtain the historical operation data of the sewage treatment system, input the pre-trained dosing prediction model, and obtain the predicted dosing amount of the lag length t i +1 i ; the predicted dosing amount and the actual dosing amount at the same time of the predicted dosing amount error minimization, and the maximum standard rate of total phosphorus in effluent as the target, select the optimal lag length t value; S3. Real-time operation parameters of the sewage treatment system are collected, and the trained dosing prediction model is used for dosing amount prediction of an optimal dosing lag time.

2. The method of dosing a sewage treatment process according to claim 1, wherein, The multi-parameter time series data set comprises a time series matrix of historical operation data from the n-tth moment to the nth moment and a historical dosing amount time series matrix covering the n+1th moment to the n+t+1th moment, and is respectively represented as: ， wherein n represents an index of a discrete time point, L represents an inflow rate, InfCOD represents an inflow COD concentration, InfAN represents an inflow ammonia nitrogen, InfTN represents an inflow total nitrogen concentration, InfTP represents an inflow total phosphorus, MLSS represents a mixed liquor suspended solids concentration of a biochemical tank, EffTP represents an effluent total phosphorus concentration, and S represents a dosing amount.

3. The method of dosing a sewage treatment according to claim 2, characterized in that, The historical operation data and the historical dosing amount are differentially processed to form time series data of a unified time scale.

4. The method of claim 3, wherein the water treatment chemicals are added to the sewage water in the form of a powder. The differential processing method is that the dosing amount S, the inflow rate L, the inflow ammonia nitrogen InfAN, the MLSS and the pH value are collected at intervals of minutes to form original minute-level time series; the inflow total nitrogen concentration InfTN, the inflow total phosphorus concentration InfTP and the inflow COD concentration InfCOD are collected at intervals of hours, and adjacent collection points are filled by a linear interpolation method to fill the time gap, and are converted into minute-level time series; the effluent total phosphorus concentration EffTP is calculated as an average of the collected values in each hour, and the hourly average of EffTP is mapped to all minute-level time nodes in the hour to generate a continuous minute-level time series.

5. The method of claim 1, wherein the method further comprises: The candidate value t i The generating method is: selecting several discrete candidate t values in the interval [t0-Δt, t0+Δt], which is represented as: t i =t0±kT, k=0,1,2,…,K, T represents a collection time step, K represents a total number of collections, kT≤Δt, i represents the i th candidate t value, i=0,1,2,…,K, and Δt represents a time offset.

6. The method of claim 1, wherein the method further comprises: By constructing a reward function, for each candidate value t i The corresponding predicted amount of reagent i And the predicted amount of reagent The actual amount of reagent S at the same time i Error, total phosphorus in effluent standard rate evaluation, by maximizing the reward function to select the optimal lag length t value, the reward function is expressed as: ， wherein, is the lag time t i the total phosphorus discharge under the standard rate, and a and β are weight coefficients.

7. A machine learning based dosing system for wastewater treatment, characterized in that, The method comprises the following modules: A multi-parameter time series data set construction module is configured to collect a plurality of groups of operation data of the sewage treatment system at preset time intervals in a historical time period, and to obtain historical dosing amount data of a lag time t+1 corresponding to each group of operation data, to construct a multi-parameter time series data set; A model construction module is configured to construct a dosing prediction model based on a random forest network, and to use the multi-parameter time series data set for training; t value updating module: taking the physical residence time t0 of sewage in the sewage treatment system as the initial t value, generating several groups of candidate values within a preset time window based on t0, for each candidate value t i , obtaining historical operation data of the sewage treatment system, inputting the pre-trained dosing prediction model, and obtaining the predicted dosing amount of the lag time t i +1 i ; the corresponding predicted dosing amount and the actual dosing amount at the same moment error minimization and maximum effluent total phosphorus standard rate as the target, the optimal lag time t value is selected; A dosing amount real-time prediction module is configured to collect real-time operation parameters of the sewage treatment system, and to use the trained dosing prediction model to predict a dosing amount of an optimal dosing lag time.

8. An electronic device, comprising: The electronic device comprises at least one processor and a memory in communication connection with the at least one processor; The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the machine learning-based sewage treatment dosing method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing the processor to implement the machine learning-based dosing method for sewage treatment according to any one of claims 1-6 when executed.

10. A computer program product, characterised in that, The computer program product comprises a computer program which, when executed by the processor, implements the machine learning-based dosing method for sewage treatment according to any one of claims 1-6.

Images