CN-121256506-B - Water quality prediction method for water treatment plant based on neural network algorithm

CN121256506BCN 121256506 BCN121256506 BCN 121256506BCN-121256506-B

Abstract

The invention relates to the technical field of water quality prediction, in particular to a water quality prediction method of a water treatment plant based on a neural network algorithm. The method comprises the steps of collecting operation data and water quality data of a water treatment plant under different working conditions in real time, preprocessing the operation data and the water quality data, extracting multidimensional features reflecting the chemical oxygen demand change rules through dynamic feature extraction, hidden state coding of sludge thickness of a sedimentation tank, static feature screening and interactive feature construction based on the preprocessed operation data and the preprocessed water quality data, predicting the water quality data based on the multidimensional features by utilizing a long-period memory network model based on a double-flow self-adaptive attention mechanism, and outputting a predicted value of the chemical oxygen demand in a future period. The invention comprehensively considers the dynamic change characteristics, the static correlation screening characteristics and the interactive coupling characteristics among the parameters of the operation parameters in the characteristic extraction stage.

Inventors

QI QINGSONG
TANG HAIYUN
HE SONG
FU JIANKANG

Assignees

广东科创智水科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251115

Claims (7)

1. The water quality prediction method of the water treatment plant based on the neural network algorithm is characterized by comprising the following steps of: S1, acquiring operation data and water quality data of a water treatment plant under different working conditions in real time, and preprocessing the operation data and the water quality data; s2, extracting multidimensional features reflecting chemical oxygen demand change rules through dynamic feature extraction, hidden state coding of sludge thickness of a sedimentation tank, static feature screening and interactive feature construction based on the pretreated operation data and water quality data; s3, based on multidimensional characteristics, predicting water quality data by using a long-short-period memory network model based on a double-flow self-adaptive attention mechanism, and outputting a predicted value of chemical oxygen demand in a future period; S4, dynamically regulating and controlling technological parameters of the water treatment plant based on the predicted value of the chemical oxygen demand, and reducing pollution emission; in the step S2, the multidimensional feature reflecting the chemical oxygen demand change rule is extracted, and the method comprises the following steps: S2.1, performing time sequence division on operation data, setting time windows, and calculating dynamic characteristics of each operation parameter in each window; s2.2, extracting thickness dynamic mode characteristics according to the sludge thickness of the sedimentation tank by using a sludge sedimentation dynamic coding method based on a sludge thickness simulator; s2.3, based on dynamic characteristics, screening input variables related to chemical oxygen demand by using a Pearson correlation coefficient as static characteristics, and constructing interaction characteristics; S2.4, carrying out normalization processing on the thickness dynamic mode characteristics, the dynamic characteristics, the static characteristics and the interaction characteristics by adopting a Z-score method; S2.5, splicing the normalized thickness dynamic modal characteristics, dynamic characteristics, static characteristics and interactive characteristics into a multidimensional characteristic vector according to time sequence ; In the step S2.2, aiming at the sludge thickness of the sedimentation tank, a sludge sedimentation dynamic coding method based on a sludge thickness simulator is utilized to extract thickness dynamic mode characteristics, and the method comprises the following steps: S2.21, calculating virtual net sludge flux at each moment according to the concentration of suspended solids in the inlet water, the inlet water flow, the concentration and the reflux flow of suspended solids in the reflux sludge, the sludge discharge amount, the concentration and the outlet water flow of suspended solids in the outlet water ; S2.22, constructing a lightweight cyclic neural network as a sludge thickness simulator, and generating a hidden state by parallel operation with a long-term and short-term memory network model based on a double-flow self-adaptive attention mechanism From the hidden state Output the estimated value of the suspension solid concentration of the water through the full connecting layer And an estimate of the chemical oxygen demand of the effluent And suspending the solid concentration estimation value through the effluent And an estimate of the chemical oxygen demand of the effluent Performing joint training on the auxiliary loss function of the model (C); S2.23, after training is completed, extracting a hidden state vector generated by the sludge thickness simulator in each time step to be used as a thickness dynamic mode characteristic of the sludge thickness; In the step S3, the water quality data is predicted by using a long-short-period memory network model based on a double-flow self-adaptive attention mechanism, and a predicted value of chemical oxygen demand in a future period is output, and the method comprises the following steps: S3.1, multidimensional feature vector Arranged in time order to construct an input sequence And to input sequence Carrying out normalization treatment; s3.2, constructing a long-period memory network model based on a double-flow self-adaptive attention mechanism; S3.3, training a long-period memory network model based on a double-flow self-adaptive attention mechanism by adopting a mean square error; s3.4, input sequence Inputting the data into a trained long-period and short-period memory network model based on a double-flow self-adaptive attention mechanism; S3.5, output future Predicted values of chemical oxygen demand for each time step.
2. The method for predicting water quality of a water treatment plant based on a neural network algorithm according to claim 1, wherein in S1, the operation data at least comprises inflow water flow, dosing amount, aeration intensity and sludge discharge amount, and the water quality data at least comprises inflow water temperature, historical outflow water chemical oxygen demand concentration, suspended solid concentration in inflow water, suspended solid concentration in return sludge, suspended solid concentration in outflow water and suspended solid concentration in sludge discharge.
3. The method for predicting water quality of water treatment plant based on neural network algorithm as set forth in claim 1, wherein in S2.22, the construction of the lightweight cyclic neural network as a sludge thickness simulator involves the specific steps of simulating net sludge flux Rate of change of As the input of the sludge thickness simulator, is used for reflecting the dynamic accumulation and discharge trend of the sludge in the sedimentation tank, and the input virtual net sludge flux is calculated by time-step recursion Is encoded as a sequence of hidden states Will hide the state Explicitly defined as a representation of normalized virtual sludge thickness and training a sludge thickness simulator through an auxiliary loss function to enable hidden states The time evolution characteristics of the water treatment system can represent the dynamic response rule of the concentration of suspended solids in water and the chemical oxygen demand.
4. The method for predicting water quality of water treatment plant based on neural network algorithm as set forth in claim 1, wherein in S3.2, constructing a long-term and short-term memory network model based on a double-flow adaptive attention mechanism involves the specific steps of taking a multidimensional feature vector Separation into process steady state feature vectors And reflow dynamic feature vectors Calculating sensitivity weight of reflux dynamic characteristics at each time step of the long-short-period memory network model, expanding an input gate, a forgetting gate and an output gate of the long-short-period memory network model into a double-flow structure, and adding a reflux mutation early-warning compensation mechanism to monitor the variation amplitude of reflux proportion in real time; wherein the reflux dynamic characteristic flow at least comprises the reflux ratio instantaneous change rate, the reflux accumulation effect index and the reflux-sludge thickness coupling characteristic.
5. The method for predicting water quality of water treatment plant based on neural network algorithm according to claim 4, wherein the step of calculating sensitivity weight of reflux dynamic characteristics at each time step of long-term and short-term memory network model comprises the following steps: At each time step Hiding the state in the previous step Current reflow dynamic feature vector Normalization processing is carried out, and the hidden state of the last step is hidden Current reflow dynamic feature vector Spliced as an attention input vector Input attention to vector Performing one-layer linear mapping and nonlinear activation to obtain intermediate representation Intermediate representation And bias term Calculating an attention score and obtaining sensitivity weights by a sigmoid activation function Weighting sensitivity Multiplying the original reflow characteristics element by element to obtain weighted reflow characteristics And inputting the data into the expanded double-flow gating long-term and short-term memory network model.
6. The method for predicting water quality of water treatment plant based on neural network algorithm according to claim 4, wherein the method for expanding the input gate, the forgetting gate and the output gate of the long-term memory network model into a double-flow structure is characterized by comprising the following specific steps: Is the steady state characteristic vector of the process And reflow dynamic feature vectors Constructing independent gating channels respectively, and calculating activation values of an input gate, a forgetting gate and an output gate of the long-term memory network model respectively; the gating channel results are weighted and combined through a gating fusion unit, and sensitivity weights based on reflux dynamic characteristics are obtained Adaptively adjusting fusion weights in the weighted combination process; Respectively calculating steady state characteristic vectors of the process And reflow dynamic feature vectors Candidate memory cells in the gating channel; The gate control output of the candidate memory unit is subjected to weighted fusion, and the candidate memory unit is updated based on the output after weighted fusion, so that collaborative memory and self-adaptive update of steady state and dynamic information are realized; and outputting the hidden state at the current moment.
7. The method for predicting water quality of a water treatment plant based on a neural network algorithm according to claim 1, wherein in S4, the predicted value based on chemical oxygen demand dynamically regulates and controls the technological parameters of the water treatment plant, and the method comprises the following steps: s4.1, comparing a chemical oxygen demand predicted value predicted by a long-term and short-term memory model with a set target chemical oxygen demand interval, and calculating a predicted water chemical oxygen demand deviation; S4.2, determining a regulation strategy and generating a regulation instruction according to the change direction and the change amplitude of the predicted water chemical oxygen demand deviation; S4.3, transmitting the regulation and control instruction to an automatic control system, and driving the aeration and backflow device to execute parameter adjustment for reducing pollutant emission.

Description

Water quality prediction method for water treatment plant based on neural network algorithm Technical Field The invention relates to the technical field of water quality prediction, in particular to a water quality prediction method of a water treatment plant based on a neural network algorithm. Background The urban domestic sewage treatment process involves complex biochemical reaction, mud-water separation, substance transfer and other multiphase coupling mechanisms, and the effluent quality (chemical oxygen demand (COD)) is influenced by various factors such as inflow water flow, water quality fluctuation, temperature change, sludge state and process parameters (such as aeration intensity, dosing amount and reflux ratio) and presents remarkable nonlinearity, time variability and uncertainty. The traditional water quality prediction method mainly depends on a mechanism model (such as an Activated Sludge Model (ASM) series), but the model has a complex structure, numerous parameters and depends on accurate initial setting and continuous calibration, so that the model is difficult to stably apply in actual engineering, and a conventional statistical model (such as multiple linear regression and a support vector machine) has a limited expression capacity, so that complex dynamic association inside a system is difficult to capture, and the prediction precision is insufficient. In addition, the key process states such as the sludge thickness, the sedimentation performance and the like of the sedimentation tank directly influence the water quality of the effluent, but the system internal state perception is lost due to the lack of a reliable on-line monitoring means, so that the accurate pre-judgment of the COD change trend of the effluent is limited. Meanwhile, the sewage treatment plant is often faced with operation disturbance such as confluence impact in rainy season, abrupt change of inflow load, abnormal backflow system and the like, the existing single-flow neural network model lacks effective distinguishing capability for steady-state operation characteristics and transient disturbance characteristics, sensitive response to sudden working conditions is difficult to realize, and prediction lag or misalignment is easy to cause. More importantly, most water plants still adopt a passive control mode of 'regulation after exceeding standard', and regulation has obvious hysteresis, so that not only is stable water output up to standard difficult to ensure, but also the problems of excessive aeration, medicament waste and the like are easily caused, and unnecessary consumption of energy and resources is caused. Therefore, a water quality prediction method of a water treatment plant based on a neural network algorithm is provided. Disclosure of Invention The invention aims to provide a water quality prediction method of a water treatment plant based on a neural network algorithm, which aims to solve the problems of low prediction precision and inaccurate operation control of effluent COD (chemical oxygen demand) caused by non-linear time-varying, undetectable key state, delayed abrupt change working condition response and rough regulation and control of the water quality in the conventional sewage treatment system. In order to achieve the above purpose, the present invention provides a water quality prediction method for a water treatment plant based on a neural network algorithm, comprising the following steps: S1, acquiring operation data and water quality data of a water treatment plant under different working conditions in real time, and preprocessing the operation data and the water quality data; s2, extracting multidimensional features reflecting chemical oxygen demand change rules through dynamic feature extraction, hidden state coding of sludge thickness of a sedimentation tank, static feature screening and interactive feature construction based on the pretreated operation data and water quality data; s3, based on multidimensional characteristics, predicting water quality data by using a long-short-period memory network model based on a double-flow self-adaptive attention mechanism, and outputting a predicted value of chemical oxygen demand in a future period; S4, dynamically regulating and controlling technological parameters of the water treatment plant based on the predicted value of the chemical oxygen demand, and reducing pollution discharge. The technical scheme is further improved, and is characterized in that in the S1, the operation data at least comprise water inflow flow, dosage, aeration intensity and sludge discharge quantity, and the water quality data at least comprise water inflow temperature, historical water outflow chemical oxygen demand concentration, suspended solid concentration in water inflow, suspended solid concentration in return sludge, suspended solid concentration in water outflow and suspended solid concentration in sludge discharge. The technical scheme is further improved