An intelligent collaborative regulation method and system for coal mine underground water treatment and recharge

Abstract

This application discloses an intelligent collaborative control method and system for coal mine groundwater treatment and replenishment. The method includes: acquiring time-series data of the influent water quality of the original mine water; using a bidirectional long short-term memory network to predict the optimal influent flow rate based on the time-series data of the influent water quality; adjusting the influent flow rate of the sedimentation tank to the optimal flow rate to perform sedimentation pretreatment on the original mine water; after pretreatment, acquiring the multidimensional membrane operating parameters of the ultrafiltration membrane; using a variational autoencoder to determine the degree of ultrafiltration membrane fouling based on the parameters; dynamically adjusting the ultrafiltration membrane cleaning strategy based on the degree of fouling and performing the cleaning operation; after cleaning, acquiring groundwater hydrological data and replenishment water quality data of the target replenishment area; using a dynamic water allocation prediction model that integrates a temporal convolutional network, a gating mechanism, and a self-attention mechanism; combining the above-mentioned groundwater hydrological data and replenishment water quality data to determine the allocation scheme for each replenishment well; and replenishing the target groundwater layer with qualified effluent through the corresponding replenishment well according to the allocation scheme.

Description

Technical Field

[0001] This application relates to the field of mine water treatment technology, and relates to, but is not limited to, an intelligent collaborative control method and system for coal mine groundwater treatment and replenishment. Background Technology

[0002] Coal mining generates a large amount of mine water, which, if discharged directly, will cause serious environmental pollution and waste of water resources. Therefore, treating mine water and replenishing it to groundwater aquifers has become an important way for coal mining enterprises to achieve green mining and water resource recycling. A typical treatment and replenishment process includes pretreatment (such as sedimentation tanks), core treatment (such as ultrafiltration membrane filtration), and replenishment to groundwater aquifers.

[0003] However, the treatment and replenishment process in related technologies suffers from problems such as lagging pretreatment control, passive membrane fouling management, and a single water allocation strategy. Summary of the Invention

[0004] In view of this, embodiments of this application provide an intelligent collaborative control method and system for coal mine groundwater treatment and replenishment, which at least solves the problems of lagging control in the pretreatment stage, passive membrane fouling management, and a single water allocation strategy.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] In a first aspect, embodiments of this application provide an intelligent coordinated control method for groundwater treatment and recharge in coal mines, the method comprising:

[0007] The time-series data of the influent water quality of the raw mine water generated by coal mining is obtained. Based on the time-series data of the influent water quality, a bidirectional long short-term memory network is used to predict the optimal influent flow rate. The influent flow rate of the sedimentation tank is adjusted to the optimal influent flow rate to perform sedimentation pretreatment on the raw mine water.

[0008] After sedimentation pretreatment, multidimensional membrane operating parameters of the ultrafiltration membrane are obtained. The degree of fouling of the ultrafiltration membrane is determined based on the multidimensional membrane operating parameters using a trained variational autoencoder. The cleaning strategy of the ultrafiltration membrane is dynamically adjusted based on the degree of fouling. The cleaning operation of the ultrafiltration membrane is then performed based on the cleaning strategy.

[0009] After cleaning, groundwater hydrological data of the target replenishment area and water quality data of the replenished water that meets the standards of the ultrafiltration membrane are obtained. Using a dynamic water allocation prediction model that integrates temporal convolutional networks, gating mechanisms, and self-attention mechanisms, the allocation scheme of each replenishment well in the target replenishment area is determined based on the groundwater hydrological data and the water quality data of the replenished water. According to the allocation scheme of each replenishment well, the qualified water is replenished to the target groundwater layer through the corresponding replenishment well.

[0010] Secondly, embodiments of this application provide an intelligent collaborative control system for groundwater treatment and replenishment in coal mines. The system includes: a data sensing layer, an intelligent analysis layer, and a decision execution layer, wherein:

[0011] The data perception layer is used to acquire the time-series data of the influent water quality of the raw mine water generated by coal mining. The intelligent analysis layer is used to predict the optimal influent flow rate based on the influent water quality time-series data using a bidirectional long short-term memory network. The decision execution layer is used to adjust the influent flow rate of the sedimentation tank to the optimal influent flow rate in order to perform sedimentation pretreatment on the raw mine water.

[0012] The data sensing layer is used to acquire multidimensional membrane operating parameters of the ultrafiltration membrane after sedimentation pretreatment. The intelligent analysis layer is used to determine the degree of fouling of the ultrafiltration membrane based on the multidimensional membrane operating parameters using a trained variational autoencoder, and dynamically adjust the cleaning strategy of the ultrafiltration membrane based on the degree of fouling. The decision execution layer is used to execute the cleaning operation of the ultrafiltration membrane based on the cleaning strategy.

[0013] The data sensing layer is used to acquire groundwater hydrological data of the target replenishment area and replenishment water quality data of the ultrafiltration membrane effluent after cleaning. The intelligent analysis layer is used to determine the allocation scheme of each replenishment well in the target replenishment area based on the groundwater hydrological data and the replenishment water quality data by using a water dynamic allocation prediction model that integrates temporal convolutional networks, gating mechanisms and self-attention mechanisms. The decision execution layer is used to replenish the effluent to the target groundwater layer through the corresponding replenishment wells according to the allocation scheme of each replenishment well.

[0014] The beneficial effects of the technical solutions provided in this application include at least the following:

[0015] Based on the temporal prediction capabilities of bidirectional long short-term memory networks, the system enables advance optimization of influent flow, avoiding poor sedimentation effects caused by water quality fluctuations, improving pretreatment stability, and achieving precise and proactive control of the pretreatment stage. A variational autoencoder accurately identifies the degree of ultrafiltration membrane fouling, dynamically adapting cleaning strategies to reduce ineffective cleaning and membrane lifespan loss, lowering maintenance costs, and achieving proactive and intelligent management of membrane fouling. A model integrating temporal convolutional networks, gating mechanisms, and self-attention mechanisms enables water allocation decisions coupled with multiple hydrological and water quality factors, improving the utilization rate of replenishment water resources and adaptability to the groundwater environment, and achieving dynamic optimization of replenishment water allocation. Connecting the three stages of pretreatment, membrane treatment, and groundwater replenishment forms a collaborative system of data acquisition, model prediction, precise control, and effect feedback, significantly reducing reliance on manual intervention, improving the automation and intelligence level of the overall treatment and replenishment system, and achieving system-wide synergistic optimization. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0017] Figure 1 A schematic flowchart illustrating an intelligent collaborative control method for coal mine groundwater treatment and replenishment provided in an embodiment of this application;

[0018] Figure 2 A schematic diagram of a settlement control model framework based on attention-enhanced bidirectional LSTM provided in this application embodiment;

[0019] Figure 3 A schematic diagram of an early diagnosis model for membrane fouling based on a variational autoencoder (VAE) provided in this application embodiment;

[0020] Figure 4 A schematic diagram of a water dynamic allocation prediction model based on gated TCN and self-attention provided in this application embodiment;

[0021] Figure 5 This is a schematic diagram of the composition structure of an intelligent collaborative control system for coal mine groundwater treatment and replenishment provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0024] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0025] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0026] This application provides an intelligent coordinated control method for coal mine groundwater treatment and replenishment, applied to an electronic device. The electronic device includes, but is not limited to, mobile phones, laptops, tablets, handheld internet devices, multimedia devices, streaming media devices, mobile internet devices, wearable devices, or other types of electronic devices. The functions implemented by this method can be achieved by a processor in the electronic device calling program code. The program code can be stored in a computer storage medium; therefore, the electronic device includes at least a processor and a storage medium. The processor can be used to process the intelligent coordinated control process of coal mine groundwater treatment and replenishment, and the memory can be used to store the data required and generated during the intelligent coordinated control process of coal mine groundwater treatment and replenishment.

[0027] Figure 1 This is a flowchart illustrating an intelligent collaborative control method for coal mine groundwater treatment and recharge, provided in an embodiment of this application. It is applied to an intelligent collaborative control system for coal mine groundwater treatment and recharge. Figure 1 As shown, the method includes at least the following steps:

[0028] Step S110: Obtain the time series data of the influent water quality of the raw mine water generated by coal mining, use a bidirectional long short-term memory network to predict the optimal influent flow rate based on the influent water quality time series data, and adjust the influent flow rate of the sedimentation tank to the optimal influent flow rate to perform sedimentation pretreatment on the raw mine water.

[0029] The physical basis of the intelligent collaborative control system is a sensor network deployed at key nodes, which constitute the system's senses and collect various types of data in real time. At the coal mine water monitoring point, pH sensors, turbidity meters, total dissolved solids (TDS) sensors, suspended solids (SS) sensors, electrical conductivity (EC) sensors, and oil content analyzers (for quantitative detection of oil content in water) can be installed to collect raw mine water quality data (i.e., the influent water quality time series data). The influent water quality time series data can include pH value, total dissolved solids, turbidity, suspended solids concentration, electrical conductivity, and oil content.

[0030] Sedimentation pretreatment is the core step in water pretreatment. The sedimentation tank in the intelligent collaborative control system is used to separate and remove suspended pollutants in the raw mine water through gravity sedimentation or flocculation-assisted sedimentation, providing low-turbidity, low-suspended-solids influent for subsequent ultrafiltration, reverse osmosis and other advanced treatment units. Since the influent flow rate can affect the sedimentation effect, an influent flow meter can be installed on the influent pipe of the sedimentation tank, and an effluent water quality sensor can be installed on the effluent pipe to monitor and evaluate the sedimentation effect through influent flow rate and effluent water quality.

[0031] The central processing unit in the intelligent collaborative control system can receive sensor data from each monitoring point and run a sedimentation control model based on a Bidirectional Long Short-Term Memory (BiLSTM) network. This model has a built-in optimal influent flow prediction algorithm, which performs time-series feature learning and analysis on the influent water quality time-series data to predict the optimal influent flow rate of the sedimentation tank. Subsequently, the central processing unit sends control commands to the influent regulating valve of the sedimentation tank, using the influent regulating valve as the actuator to complete the precise adjustment of the influent flow rate, ensuring the sedimentation effect and adaptability to subsequent processes.

[0032] The influent water quality time series data can be a 6-dimensional time series feature vector with a time window length of T, and the time series feature vector X at time t in the influent water quality time series data is... t It can be expressed by the following formula (1):

[0033] X t =[xt1,xt2,...,xt6]∈R 6 Formula (1);

[0034] Where xt1 represents pH value (dimensionless), xt2 represents total dissolved solids (TDS) (mg / L), xt3 represents turbidity (NTU), xt4 represents suspended solids concentration (mg / L), xt5 represents conductivity (μS / cm), and xt6 represents oil content (mg / L); the output variable can be represented as y. t ∈R, representing the optimal influent flow rate (m³ / h) of the settling tank at time t, R 6 It represents a 6-dimensional real number space.

[0035] Step S120: After sedimentation pretreatment, obtain the multidimensional membrane operating parameters of the ultrafiltration membrane, use the trained variational autoencoder to determine the fouling degree of the ultrafiltration membrane based on the multidimensional membrane operating parameters, dynamically adjust the cleaning strategy of the ultrafiltration membrane based on the fouling degree, and perform the cleaning operation of the ultrafiltration membrane based on the cleaning strategy.

[0036] In the intelligent collaborative control system, transmembrane pressure (TMP) sensors can be installed on the feed water and product water sides of the ultrafiltration membrane, a product water turbidity meter can be installed on the ultrafiltration membrane product water main pipe, and an online microbial diversity index (such as the Shannon index) analyzer can be configured. By monitoring TMP changes, product water turbidity fluctuations, and the evolution of the microbial community structure on the membrane surface (using the Shannon index as a quantitative characterization), abnormal microbial diversity can be used as an early warning indicator of ultrafiltration membrane fouling. The multidimensional membrane operating parameters can include transmembrane pressure, product water turbidity, and microbial diversity index.

[0037] The central processing unit in the intelligent collaborative control system can construct an early diagnosis model of membrane fouling based on the multidimensional membrane operating parameters using a variational autoencoder (VAE). This model accurately identifies the degree of fouling of the ultrafiltration membrane, which can be divided into two categories: one is mild fouling, moderate fouling, and severe fouling, and the other is normal operation, mild fouling, and severe fouling. Appropriate cleaning strategies are dynamically generated according to different degrees of fouling. Subsequently, the central processing unit sends control commands to the ultrafiltration membrane cleaning device (such as a chemical cleaning pump, physical cleaning unit, etc.), and the cleaning device acts as the actuator to perform the cleaning operation.

[0038] Step S130: After cleaning, obtain the groundwater hydrological data of the target replenishment area and the water quality data of the replenishment water that meets the standards of the ultrafiltration membrane. Using a water volume dynamic allocation prediction model that integrates temporal convolutional networks, gating mechanisms and self-attention mechanisms, determine the allocation scheme of each replenishment well in the target replenishment area based on the groundwater hydrological data and the water quality data of the replenishment water. According to the allocation scheme of each replenishment well, the water that meets the standards is replenished to the target groundwater layer through the corresponding replenishment well.

[0039] Specifically, groundwater level gauges and resistivity meters can be deployed in each replenishment well and its surrounding area to form a groundwater monitoring network, thereby acquiring groundwater flow field and geological information in real time. Flow meters, TDS sensors, and online COD (Chemical Oxygen Demand) analyzers can also be installed at the inlet of each replenishment well to monitor the water quality and quantity of the replenishment water. The groundwater hydrological data includes groundwater level and resistivity, and the replenishment water quality data includes the total dissolved solids and chemical oxygen demand of the replenishment water in each of the replenishment wells.

[0040] The dynamic water allocation prediction model adopts a hierarchical progressive architecture, comprising four core components: an input embedding layer, a gated temporal convolutional network (TCN) encoder, a self-attention fusion layer, and a prediction output layer. In the intelligent collaborative control system, the central processing unit (CPU) first standardizes and aligns the multi-source input data of the input embedding layer to obtain model input data in a unified format. This input data is then processed sequentially through the dilated convolution and gating mechanism of the gated TCN encoder to extract multi-level temporal features. The self-attention fusion layer then uses its self-attention mechanism to dynamically assign weights to features of different dimensions and enhance key features. Finally, the fully connected network of the prediction output layer generates the optimal water allocation scheme for each replenishment well. Subsequently, the CPU sends control commands to the water distribution regulating valves of each replenishment well. These valves act as actuators, responding to and executing water allocation operations to ensure efficient and compliant replenishment.

[0041] In the above embodiments, the temporal prediction capability of bidirectional long short-term memory networks enables advance optimization of influent flow, avoiding poor sedimentation effects caused by water quality fluctuations, improving pretreatment stability, and achieving precise and proactive control of the pretreatment stage. Variational autoencoders accurately identify the degree of ultrafiltration membrane fouling, dynamically adapt cleaning strategies, reduce the loss of membrane life due to ineffective cleaning, lower maintenance costs, and achieve proactive and intelligent control of membrane fouling. A model integrating temporal convolutional networks, gating mechanisms, and self-attention mechanisms enables water allocation decisions coupled with multiple hydrological and water quality factors, improving the utilization rate of replenishment water resources and adaptability to the groundwater environment, and achieving dynamic optimization of replenishment water allocation. Connecting the three stages of pretreatment, membrane treatment, and groundwater replenishment forms a collaborative system of data acquisition, model prediction, precise control, and effect feedback, significantly reducing reliance on manual intervention, improving the automation and intelligence level of the overall treatment and replenishment system, and achieving system synergistic optimization.

[0042] In some embodiments, such as Figure 2As shown, the sedimentation control model includes an input layer, the bidirectional long short-term memory network, an attention mechanism layer, a triple fully connected layer, and an output layer. The bidirectional long short-term memory network includes a bidirectional LSTM encoding layer, which includes a forward LSTM encoding layer and a backward LSTM encoding layer. The triple fully connected layer is a network module composed of three fully connected layers stacked sequentially. The influent water quality time series data includes pH value, total dissolved solids, turbidity, suspended solids concentration, conductivity, and oil content.

[0043] Step S110, "predicting the optimal influent flow rate based on the influent water quality time series data using a bidirectional long short-term memory network," includes the following steps:

[0044] Step S1101: Extract the forward temporal features of the influent water quality time series data using the forward LSTM encoding layer of the bidirectional long short-term memory network, extract the backward temporal features of the influent water quality time series data using the backward LSTM encoding layer of the bidirectional long short-term memory network, and concatenate the forward and backward temporal features according to the time step dimension to obtain bidirectional concatenated temporal features, wherein the bidirectional concatenated temporal features include feature vectors corresponding to multiple time steps;

[0045] The bidirectional long short-term memory network acquires the influent water quality time-series data of the input layer. The forward time-series features can be represented by the following formula (2), the backward time-series features can be represented by the following formula (3), and the bidirectional splicing time-series features can be represented by the following formula (4):

[0046] Formula (2);

[0047] Formula (3);

[0048] Formula (4);

[0049] in, This represents the hidden state (hidden vector) output by the feedforward LSTM coding layer at time t-1. This represents the hidden state output by the forward LSTM coding layer at time t; This represents the hidden state output to the LSTM coding layer after time t-1. X represents the hidden state output to the LSTM coding layer after time t; t This represents the time-series feature vector at time t in the influent water quality time-series data; This indicates the timing characteristics of bidirectional splicing.

[0050] Step S1102: Using the attention score vector, attention weight matrix, attention bias, and the bidirectional splicing temporal features of the attention mechanism layer, determine the attention weight of each time step in the bidirectional splicing temporal features; based on the attention weight of each time step, perform weighted fusion on the feature vectors of multiple time steps in the bidirectional splicing temporal features to obtain a weighted feature vector.

[0051] The attention mechanism layer assigns weights to the bidirectional splicing time-series features, focusing on the features at key time points, thereby improving the settlement control model's ability to capture key information and enhancing its interpretability. The attention score of each time step in the bidirectional splicing time-series features can be calculated using the following formula (5):

[0052] Formula (5);

[0053] in, This represents the attention score at time step t in the bidirectional splicing temporal features. Represents the attention score vector. Represents the attention weight matrix. This indicates attentional bias.

[0054] The attention weights of each time step in the bidirectional splicing temporal features can be calculated using the following formula (6) based on the attention scores at each time step. :

[0055] Formula (6);

[0056] in, This represents the attention score at time step j in the bidirectional splicing temporal features, where j ranges from 1 to T, and T represents the total length of the time window. The attention weight at time t represents the attention weight, which can be dynamically allocated based on the attention mechanism. It focuses on time points with abnormal fluctuations and reflects the importance of water quality changes at each time point to decision-making.

[0057] The weighted feature vector (i.e., the context vector) c can be calculated using the following formula (7):

[0058] Formula (7);

[0059] Step S1103: Using the fully connected weight matrix, fully connected bias, and the weighted eigenvector of the output layer, the initial influent flow rate is determined through linear transformation and activation function calculation.

[0060] It should be noted that the weighted feature vector c output by the attention mechanism layer can also be subjected to multi-layer nonlinear transformation and dimension mapping through a triple fully connected layer to enhance the nonlinear fitting ability of the settlement control model and prevent overfitting. The transformed weighted feature vector will be used as input and passed to the final output layer. The output layer performs nonlinear transformation and dimension mapping on the transformed weighted feature vector output by the triple fully connected layer to finally output the initial influent flow rate, which can be calculated by the following formula (8):

[0061] Formula (8);

[0062] The fully connected layer weight matrix includes a triple fully connected layer weight matrix and an output layer weight matrix, and the fully connected layer bias includes a triple fully connected layer bias and an output layer bias. Indicates the initial influent flow rate. This represents the weight matrix of the triple fully connected layer. This indicates the bias of the triple fully connected layer. This represents the output layer weight matrix. This indicates the output layer bias.

[0063] Step S1104: Determine the optimal influent flow rate based on the initial influent flow rate, the suspended solids concentration, the preset baseline flow rate, and the preset suspended solids safety threshold.

[0064] In some embodiments, step S1104 includes the following steps:

[0065] Step S11041: When the suspended solids concentration is greater than the preset suspended solids safety threshold, the smaller value among the initial influent flow rate, the preset reference flow rate, and the preset proportional coefficient is determined as the optimal influent flow rate, where the preset proportional coefficient is less than 1.

[0066] The initial influent flow rate can also be expressed as Q. pred The suspended solids concentration can be expressed as SS. current The preset baseline flow rate can be expressed as Q. base The preset suspended matter safety threshold can be expressed as SS. threshold The optimal influent flow rate can be expressed as: The optimal influent flow rate can be expressed by the following formula (9):

[0067] ;

[0068] Wherein, as shown in formula (9), the preset proportional coefficient can be 0.7, in At that time, and The smaller of the product of 0.7 and 0.7 is determined as the optimal influent flow rate. .

[0069] Step S11042: When the suspended solids concentration is less than or equal to the preset suspended solids safety threshold and the initial influent flow rate is greater than or equal to the preset reference flow rate, the larger of the initial influent flow rate and the preset reference flow rate is determined as the optimal influent flow rate.

[0070] As shown in formula (9), in At that time, The larger value among them is determined as the optimal influent flow rate. .

[0071] Step S11043: When the suspended solids concentration is less than or equal to the preset suspended solids safety threshold and the initial influent flow rate is less than the preset reference flow rate, the initial influent flow rate is determined as the optimal influent flow rate.

[0072] As shown in formula (9), in When (i.e., the "otherwise" case in formula (9), directly use Determined as the optimal influent flow rate .

[0073] In the above embodiments, based on the comparison between suspended solids concentration and safety threshold, a scenario-specific flow adjustment logic is established to achieve dynamic response to water quality fluctuations and flow adaptation, avoiding poor sedimentation effect caused by a single flow strategy; for scenarios with excessive suspended solids, a flow restriction mechanism of "baseline flow rate × proportional coefficient" is set to effectively avoid overloading of the sedimentation tank under high suspended solids load and ensure stable effluent water quality; for scenarios with compliant suspended solids, the upper / lower flow limit constraints are optimized to maximize treatment efficiency while ensuring sedimentation effect, balance water quality safety and treatment capacity, and improve the flexibility and practicality of the pretreatment process.

[0074] In one embodiment, the model decision rule includes three layers of decision priority logic, corresponding to traffic control strategies in different scenarios. The higher the level, the stronger the priority.

[0075] In the first stage, model prediction is dominant, and when water quality parameters are normal (SS) current ≤SS threshold ), fully trust the model's predicted values; confidence assessment based on attention weights: if max( If the value is greater than 0.3, the prediction is considered reliable.

[0076] In the second level, threshold protection is triggered when the measured SS exceeds the standard ( When the load exceeds the limit, immediately activate protective load reduction. The load reduction range is dynamically adjusted according to the degree of exceedance. The load reduction range (the amount of flow reduction) ΔQ can be calculated using the following formula (10):

[0077] ;

[0078] In the third level, expert rules serve as a fallback, switching to rule control based on expert experience when the model output is abnormal or the confidence level is insufficient.

[0079] In response to the drawback of traditional settlement control methods being ineffective due to response lag, this application proposes an advanced control model based on attention-enhanced bidirectional LSTM. This model consists of three core components: a bidirectional LSTM encoding layer, an attention mechanism layer, and a fully connected prediction layer. The bidirectional LSTM encoding layer is used to extract forward and backward time-series features, the attention mechanism layer is used to dynamically calculate the importance weights of each time step, and the fully connected prediction layer is used to output flow prediction values ​​based on weighted features.

[0080] The core principle of this model is to achieve a paradigm shift from passive response to proactive prediction. It first utilizes a bidirectional long short-term memory network to deeply process historical time-series data of influent water quality. This network comprehensively captures the complete dynamic characteristics of water quality parameter changes by combining forward-to-back causal flow and backward-to-forward contextual flow. Building on this, the model introduces a dynamic attention mechanism, which automatically evaluates and assigns differentiated importance weights to different time points in the sequence, thereby accurately identifying key change patterns such as pollution load peaks and filtering out irrelevant noise. Finally, the model uses this time-series feature expression, which incorporates key information, to predict the state of the sedimentation system in the future and generates optimal control commands such as influent flow rate or chemical dosing, thus achieving proactive regulation, fundamentally overcoming hysteresis, and improving the stability and operational efficiency of the sedimentation process.

[0081] In the above embodiments, the bidirectional LSTM model based on the attention mechanism enables advanced prediction and precise control of influent water quality, significantly improving the settling unit's resistance to shock loads. Operational results show that the system can adjust the influent flow rate before suspended solids (SS) concentration exceeds the standard, effectively ensuring the stability of the effluent water quality and providing stable influent conditions for subsequent membrane treatment processes. This significantly improves the stability and shock resistance of the treatment system.

[0082] In the above embodiments, the core dimensions of the influent water quality time series data (pH, TDS, turbidity, etc.) are clearly defined to ensure the comprehensiveness and relevance of the data input, providing a reliable data foundation for flow prediction. A bidirectional LSTM encoding layer is used to simultaneously extract forward and backward time series features, fully capturing the historical variation patterns and future trend correlations of the water quality data, solving the problem of incomplete time series feature extraction in traditional unidirectional models. An attention mechanism layer is introduced to weightedly fuse the bidirectional spliced ​​features, automatically focusing on the key time step features for flow regulation, suppressing noise interference, and improving the accuracy of flow prediction. Through linear transformation and activation function calculation in a fully connected layer, the initial influent flow rate is quantitatively output, providing a scientific basis for subsequent optimal flow rate determination and supporting proactive regulation in the pretreatment stage.

[0083] In some embodiments, such as Figure 3 As shown, the early diagnosis model for membrane fouling includes an input layer, a variational autoencoder, and an output layer. The multidimensional membrane operating parameters include transmembrane pressure difference, permeate turbidity, and microbial diversity index. Step S120, "determining the degree of fouling of the ultrafiltration membrane based on the multidimensional membrane operating parameters using the trained variational autoencoder," includes:

[0084] Step S1201: Using the three-layer fully connected network of the probabilistic encoder of the trained variational autoencoder, depth features are extracted from the multidimensional membrane operating parameters, and the mean and variance of the posterior Gaussian distribution of the latent space are output.

[0085] Among them, such as Figure 3 As shown, the trained variational autoencoder adopts an encoder-decoder architecture, consisting of three core components (probabilistic encoder, probabilistic decoder, and adaptive prior network) to form a complete probability generation network. First, the probabilistic encoder (inference network) receives the multidimensional membrane running parameters as input, extracts features through a three-layer fully connected network, and outputs the posterior distribution parameters (mean and variance) of the latent space. Each layer of the network includes batch normalization and Dropout operations to improve training stability.

[0086] Step S1202: Using the reparameterization sampling module of the variational autoencoder, latent variables are obtained through reparameterization sampling based on the mean and the variance; using a three-layer fully connected network in the probabilistic decoder of the variational autoencoder that is symmetrical to the three-layer fully connected network of the probabilistic encoder, the latent variables are reconstructed to obtain the reconstructed data of the multidimensional membrane operating parameters.

[0087] The latent variable can be represented as z. To achieve backpropagation, the reparameterization technique shown in formula (11) can be used to transform the random variable (i.e., the latent variable z) into a deterministic function:

[0088] Formula (11);

[0089] Where, μ z , σ z This represents the mean and standard deviation of the posterior distribution of the probability encoder output. This represents random noise sampled from a standard Gaussian distribution, and ⊙ represents element-wise multiplication. Represents random noise It follows a standard normal distribution with a mean of 0 and a variance of 1.

[0090] Subsequently, the probabilistic decoder (generator network) receives the latent variable z sampled from the latent space. Through a three-layer fully connected network symmetric to the probabilistic encoder, it reconstructs the probability distribution corresponding to the input data. Sampling based on this probability distribution yields reconstructed data with the same dimension as the input data, thus restoring the original input data. This model introduces an adaptive prior network to learn the prior distribution parameters (μ_prior and σ_prior) unique to the ultrafiltration membrane system, replacing the standard Gaussian prior of traditional VAEs, thereby better adapting to the actual operating characteristics of the membrane system. The entire data flow follows a closed-loop process: real-time sensor data, feature extraction, probabilistic encoder, latent space sampling, probabilistic decoder, and data reconstruction, forming a complete probabilistic inference framework.

[0091] Step S1203: Using the output layer of the variational autoencoder, based on the reconstructed data and the multidimensional membrane operating parameters, determine the reconstruction probability of the multidimensional membrane operating parameters; determine the relative entropy between the approximate posterior distribution of the latent variable and the prior distribution of the latent variable, and determine the relative entropy as the latent space deviation.

[0092] The reconstruction probability can be calculated using the following formula (12):

[0093] Formula (12);

[0094] Where, p recon The value represents the reconstruction probability, and x represents the input data vector, i.e., the real-time multidimensional membrane operating parameters. σ represents the reconstructed data, i.e., the output data of the model reconstruction; σ represents the standard deviation of the Gaussian distribution, which is used to control the sensitivity of the reconstruction probability.

[0095] The potential spatial deviation can be calculated using the following formula (13):

[0096] Formula (13);

[0097] in, This indicates that the latent variable represents an approximate posterior distribution based on the input data vector x. This represents the prior distribution of the latent variable. This represents the KL divergence operation, also known as the relative entropy operation.

[0098] Step S1204: Determine the anomaly score based on the reconstruction probability, the latent space deviation, the preset weight coefficient, and the preset maximum latent space deviation;

[0099] The abnormal score can be determined using the following formula (14):

[0100] Formula (14);

[0101] in, This represents the preset weighting coefficient, used to balance the weights of the two indicators (reconstruction probability and latent spatial deviation); This indicates the preset maximum potential spatial deviation.

[0102] Step S1205: Determine the degree of fouling of the ultrafiltration membrane based on the abnormal score, the preset normal period mean, and the preset normal period standard deviation.

[0103] In some embodiments, step S1205 includes the following steps:

[0104] Step S12051: Determine the normal threshold and the warning threshold based on the preset normal period mean and the preset normal period standard deviation;

[0105] The normal threshold can be determined using the following formula (15):

[0106] Formula (15);

[0107] Where, θ normal This represents the normal threshold, i.e., the upper boundary threshold of the normal state; μ normal σ represents the preset normal period mean, i.e., the average of the abnormal scores during historical normal periods; normal This represents the standard deviation of the preset normal period, which is the standard deviation of the abnormal scores in the historical normal period.

[0108] The warning threshold can also be determined using the following formula (16):

[0109] Formula (16);

[0110] Where, θ warning This indicates the warning threshold, which is the lower boundary threshold that triggers a critical alarm.

[0111] Step S12052: If the abnormal score is less than the normal threshold, determine that the fouling level of the ultrafiltration membrane is normal operation.

[0112] The fouling level of the ultrafiltration membrane includes Normal Operation, Minor Fouling Alert, and Severe Fouling Alarm; as shown in the following formula (17), the fouling level of the ultrafiltration membrane can be determined by comparing the anomaly score with the normal threshold and the warning threshold:

[0113] Formula (17);

[0114] As shown in formula (17) above, in Less than θ normal Under these circumstances, the degree of fouling of the ultrafiltration membrane is determined to be normal operating condition.

[0115] Step S12053: If the abnormal score is greater than or equal to the normal threshold and less than the warning threshold, the fouling degree of the ultrafiltration membrane is determined to be a mild fouling warning.

[0116] As shown in formula (17) above, in And less than θ warning In this case, the degree of contamination of the ultrafiltration membrane is determined to be a mild contamination warning.

[0117] Step S12054: If the abnormal score is greater than or equal to the warning threshold, determine that the fouling degree of the ultrafiltration membrane is a severe fouling warning.

[0118] As shown in formula (17) above, in Greater than or equal to θ warning In this case, the degree of contamination of the ultrafiltration membrane is determined to be a severe contamination warning.

[0119] In the above embodiments, the logic for dividing normal thresholds and warning thresholds is clearly defined, and a three-level control system of "normal operation - mild pollution warning - severe pollution warning" is established to make membrane fouling status visible and quantifiable. Different treatment strategies are corresponding to different pollution levels to avoid the problems of excessive cleaning for mild pollution or untimely cleaning for severe pollution, thereby extending membrane life and reducing operation and maintenance costs. The warning mechanism avoids the risk of membrane performance degradation in advance, ensures the stability of ultrafiltration membrane treatment efficiency, and provides support for subsequent water quality compliance.

[0120] This model employs a multi-level, adaptive intelligent decision-making mechanism, combining quantitative assessment, dynamic thresholds, and trend prediction to achieve accurate judgment and early warning of membrane fouling status. The decision-making process first calculates the reconstruction probability and potential spatial distribution deviation of real-time data to generate a comprehensive anomaly score; then, it dynamically adjusts the decision threshold based on the statistical characteristics of historical data to achieve adaptive classification; finally, it ensures the accuracy and reliability of the decision through trend prediction and confidence assessment.

[0121] This model, based on variational inference theory, transforms the problem of early membrane fouling diagnosis into an anomaly detection task. The core idea is to establish a baseline model of the membrane system's health status by learning the probability distribution characteristics of operating data under normal conditions. When real-time operating data deviates significantly from the learned normal distribution, it can be identified as an early sign of fouling. Compared to traditional threshold methods, this approach has stronger generalization and early warning capabilities, issuing warnings at the initial stage of fouling (typically 0.5 to 2 hours earlier than traditional methods).

[0122] In the above embodiments, the early diagnostic model based on a variational autoencoder (VAE) enables quantitative identification of membrane fouling, significantly improving diagnostic accuracy. Compared to traditional timed cleaning methods, it significantly optimizes the cleaning cycle, reduces the incidence of irreversible membrane fouling, and is expected to effectively extend membrane life, reduce cleaning agent consumption, and decrease unplanned downtime. This significantly reduces operating and maintenance costs and extends the lifespan of critical equipment.

[0123] In the above embodiments, transmembrane pressure difference, permeate turbidity, and microbial diversity index are selected as core operating parameters to comprehensively cover the physical, chemical, and biological characteristics of membrane fouling, solving the problem of incomplete fouling judgment by traditional single parameters. A VAE probabilistic encoder / decoder is used to achieve deep feature extraction and reconstruction of multidimensional parameters. Combined with reconstruction probability and potential spatial deviation to quantify parameter anomalies, the sensitivity and accuracy of fouling identification are improved. A fouling degree grading standard is established based on anomaly scores and the mean / standard deviation of the normal period, enabling quantitative judgment of the fouling state (rather than qualitative description), providing a precise basis for subsequent dynamic cleaning strategies, and transforming passive cleaning into proactive management.

[0124] In some embodiments, the method further includes:

[0125] Step S101: Input the historical multidimensional membrane running parameters in the training dataset into the initial variational autoencoder to be trained to obtain the corresponding historical reconstruction data.

[0126] The historical multidimensional membrane operating parameters can be expressed as follows: The historical reconstructed data can be represented as .

[0127] Step S102: Based on the historical reconstruction data and the historical multidimensional membrane operating parameters, determine the reconstruction loss during the training process;

[0128] The reconstruction loss during training can be determined using the following formula (18):

[0129] Formula (18);

[0130] in, Indicates the reconstruction loss. It is the mean square error. It is the structural similarity index, and λ is a hyperparameter that balances the weights of the two terms (mean squared error and structural similarity index).

[0131] Step S103: Based on the mean and variance of the initial variational autoencoder output, and the preset prior mean and prior variance, determine the relative entropy loss during the training process.

[0132] The relative entropy loss during training can be determined using the following formula (19):

[0133] Formula (19);

[0134] in, This represents the relative entropy loss. Indicates a normal distribution. This represents the mean and variance of the output of the initial variational autoencoder. This represents the preset prior mean and prior variance (parameters of the learnable prior distribution).

[0135] Step S104: Based on the temporal sequence corresponding to the latent variables output by the initial variational autoencoder, determine the temporal consistency loss during the training process;

[0136] The temporal consistency loss during training can be determined using the following formula (20):

[0137] Formula (20);

[0138] in, Indicates the time-series consistency loss. This represents the latent vector corresponding to the t-th time step. This represents the latent vector corresponding to the (t-1)th time step. This represents the total number of time steps.

[0139] Step S105: Based on the reconstruction loss, the relative entropy loss, and the temporal consistency loss, the value of the multi-objective loss function is determined by weighted fusion using preset weight coefficients;

[0140] As shown in formula (21) below, the model adopts an improved multi-objective loss function, which comprehensively improves the early diagnosis performance of membrane fouling by jointly optimizing reconstruction accuracy, latent spatial regularity and temporal consistency:

[0141] Formula (21);

[0142] in, This represents the value of the multi-objective loss function. and This is a hyperparameter used to balance the weights of different losses.

[0143] Step S106: Based on the value of the multi-objective loss function, the initial variational autoencoder is iteratively updated using the gradient descent method until the value of the multi-objective loss function converges, thus obtaining the trained variational autoencoder.

[0144] In the above embodiments, a multi-objective loss function is constructed by introducing reconstruction loss, relative entropy loss, and temporal consistency loss, which takes into account the accuracy of feature reconstruction, the rationality of potential spatial distribution, and the continuity of temporal data, thereby improving the stability and generalization ability of model training. The model parameters are iteratively optimized by gradient descent to ensure that the VAE can accurately learn the nonlinear mapping relationship between multidimensional membrane operating parameters and pollution level, solving the problems of poor generalization ability and easy overfitting of traditional models. The trained VAE model has reliable pollution identification ability, providing core technical support for the proactive management of membrane pollution and ensuring the long-term stable operation of membrane treatment.

[0145] In some embodiments, such as Figure 4 As shown, the dynamic water allocation prediction model includes an input layer, an enhanced temporal convolutional network layer, and an output layer. The enhanced temporal convolutional network layer integrates a temporal convolutional network, a gating mechanism, and a self-attention mechanism. The temporal convolutional network includes multiple dilated causal convolutional layers. The groundwater hydrological data includes groundwater level and resistivity. The recharge water quality data includes the total dissolved solids and chemical oxygen demand of the recharge water for each recharge well. Step S130, "Using the dynamic water allocation prediction model that integrates a temporal convolutional network, a gating mechanism, and a self-attention mechanism, based on the groundwater hydrological data and the recharge water quality data, determine the allocation scheme for each recharge well within the target recharge area," includes:

[0146] Step S1301: Based on the dilation factor, kernel size, and kernel weight parameters of each dilated causal convolutional layer in the multi-layer dilated causal convolutional network, long-term causal features are extracted from the groundwater data by sampling time-series data at intervals.

[0147] Among them, the water dynamic allocation prediction model integrates dilated causal convolution and gating mechanism to enhance feature selection ability while maintaining temporal causal relationship. Specifically, feature mapping calculation and information flow regulation are achieved through the following formula (22):

[0148] Formula (22);

[0149] Among them, W (l) (k) is the weight parameter of the k-th convolutional kernel in the l-th layer, d is the dilation factor (representing sampling at intervals of d time steps), and k is the size of the convolutional kernel, which controls the range of local feature extraction. No. The output features of the layer at time t. No. Layer, First Time (i.e., before) The input features at time (time). Formula (22) controls the interval by using the dilation factor d to control the interval, the convolution kernel k to control the local range, and the weight parameter to control the feature mapping, thus realizing the process of sampling historical data at intervals and extracting long-term causal features.

[0150] Step S1302: Based on the feature transformation convolution kernel weight matrix, the long-term causal features are transformed to obtain transformed features. Based on the gated convolution kernel weight matrix and the long-term causal features, a gated signal is generated. Based on the transformed features and the gated signal, key features strongly correlated with membrane fouling are screened out through the gated mechanism.

[0151] Key features strongly correlated with membrane fouling can be screened using the following formula (23):

[0152] Formula (23);

[0153] in, It is a key feature. W conv W is the feature transformation convolution kernel weight matrix. gate It is the gated convolution kernel weight matrix, used to control the information flow rate; Indicates the transformed features. This represents a gating signal, and ⊙ represents element-wise multiplication to achieve gating control.

[0154] Step S1303: Based on the adaptive weight matrix, weight projection vector, and weight parameters of the convolution kernel, determine the adaptive importance weight of each convolution kernel; based on the adaptive importance weight of each convolution kernel, filter the convolution kernels to obtain high-importance convolution kernels; based on the high-importance convolution kernels, extract features from the key features to obtain high-quality feature vectors.

[0155] The adaptive importance weights of each convolutional kernel can be determined using the following formula (24):

[0156] Formula (24);

[0157] Where, α k W is the adaptive importance weight of the k-th convolutional kernel; v is the weight projection vector used to map high-dimensional features to weight scores; adapt It is an adaptive weight matrix used to learn feature importance; W conv (k) is the feature transformation convolution kernel weight matrix W. conv The weight parameters of the k-th convolutional kernel.

[0158] After obtaining the adaptive importance weights for each convolutional kernel, kernels with high weights can be retained, while those with low weights can be discarded or weakened. Key features are then extracted using the high-importance weights to obtain a high-quality feature vector. This high-quality feature vector may include the temporal features h at time step i. i and the temporal characteristics h at time j j It contains time-series characteristic information.

[0159] Step S1304: Based on the query weight matrix and the key weight matrix, map any two high-quality feature vectors to query vectors and key vectors respectively; based on the inner product of the query vector and the key vector and the square root of the dimension of the key vector, obtain the cross-attention weights between temporal features.

[0160] The cross-attention weights between temporal features can be obtained using the following formula (25):

[0161] Formula (25);

[0162] in, It is the cross-attention weight, W Q W K These are the query weight matrix and the key weight matrix, used to calculate feature similarity; d k It is the dimension of the key vector, used to scale the cross-attention score. .

[0163] Step S1305: Based on the value weight matrix, the high-quality feature vector at each time step is converted into a value vector. Based on the cross attention weight, the value vectors at multiple time steps are weighted and fused to obtain the attention-weighted fusion feature corresponding to each backfill well.

[0164] The attention-weighted fusion features corresponding to each replenishment well can be obtained through the following formula (26):

[0165] Formula (26);

[0166] Among them, z i It is the attention-weighted fusion feature corresponding to the i-th well, T is the length of the input feature sequence (i.e., the total number of time steps of the time sequence features), which determines the scope of attention calculation, W V It is a value weight matrix used to weight time series features h j Perform dimensional transformation and representation optimization.

[0167] Step S1306: Based on the attention-weighted fusion features, hidden layer weight matrix and hidden layer bias vector, output layer weight matrix and output layer bias vector corresponding to each replenishment well, determine the total water volume prediction value of the target replenishment area;

[0168] The set of attention-weighted fusion features corresponding to all replenishment wells can be represented as f, and the total water volume prediction value of the target replenishment area can be determined by the following formula (27):

[0169] Formula (27);

[0170] in, This is the predicted total water volume, W. hid b hid These are the hidden layer weight matrix and the hidden layer bias vector; W out b out These are the output layer weight matrix and the output layer bias vector.

[0171] Step S1307: Based on the groundwater hydrological data and hydrological response function, determine the hydrological response function value; based on the recharge water quality data and water quality adjustment function, determine the water quality adjustment function value.

[0172] Step S1308: Based on the total water volume prediction value, the water allocation weight parameter of each replenishment well, the hydrological response function value, and the water quality adjustment function value, the total water volume prediction value is allocated to each replenishment well to obtain the initial water allocation of the corresponding replenishment well. The water allocation weight parameter of each replenishment well is dynamically updated based on the replenishment efficiency of the corresponding replenishment well.

[0173] It should be noted that the process of determining the total water volume prediction value based on the groundwater hydrological data using the water volume dynamic allocation prediction model that integrates temporal convolutional networks, gating mechanisms, and self-attention mechanisms in steps S1301 to S1306 above can also be expressed by the following formula (28):

[0174] Formula (28);

[0175] in, This represents a dynamic water allocation prediction model. This indicates the current hydrological parameters (such as water level and resistivity) in the groundwater hydrological data. This refers to the historical time-series data in the aforementioned groundwater hydrological data.

[0176] The initial water distribution for each replenishment well can be determined using the following formulas (29) and (30):

[0177] Formula (29);

[0178] Formula (30);

[0179] in, This represents the water distribution reference value for the i-th replenishment well based on water quality constraints. w represents the water distribution reference value for the i-th replenishment well based on hydrological constraints. i f is the water distribution weighting coefficient for the i-th well location. TDS and f COD For water quality adjustment function, g level and g resistivity For hydrological response function, This represents the TDS value of the i-th replenishment well. This represents the COD value of the i-th replenishment well. This represents the water level of the i-th replenishment well. This represents the resistance of the i-th replenishment well. Similarly, This represents the TDS value of the j-th replenishment well. This represents the COD value of the j-th replenishment well. This represents the water level of the j-th replenishment well. This represents the resistivity of the j-th replenishment well.

[0180] Based on and By performing weighted fusion, the initial water distribution volume of the i-th replenishment well is obtained. .

[0181] It should be noted that, as shown in the following formula (31), the water distribution weight parameter of each replenishment well is dynamically updated based on the replenishment efficiency of the corresponding replenishment well:

[0182] Formula (31);

[0183] in, It is the partial derivative of the replenishment efficiency with respect to the water distribution weight parameter, and η is the learning rate, which controls the update speed of the water distribution weight parameter. This represents the water allocation weight parameter for replenishment well i at the current time (or in the current round). This represents the water distribution weight parameter for replenishment well i at the next moment (or in the next round).

[0184] Step S1309: Based on the multi-constraint optimization strategy, the initial water distribution of each replenishment well is optimized to obtain the target allocation scheme for the corresponding replenishment well.

[0185] In some embodiments, step S1309, "based on a multi-constraint optimization strategy, optimizing the initial water distribution of each replenishment well to obtain the target allocation scheme for the corresponding replenishment well," includes the following steps:

[0186] Step S13091: If the total dissolved solids in the replenishment water of any of the replenishment wells are greater than a preset total dissolved solids safety threshold, determine the change in the total dissolved solids in the replenishment water of the corresponding replenishment well; if the change is greater than or equal to the change threshold, adjust the initial water distribution of the corresponding replenishment well based on the change in the total dissolved solids in the replenishment water of each replenishment well and the emergency attenuation coefficient to obtain the emergency water distribution of the corresponding replenishment well; determine the smaller value between the emergency water distribution of the corresponding replenishment well and the preset maximum allowable water distribution as the candidate water distribution of the corresponding replenishment well.

[0187] The emergency water distribution volume of the replenishment well can be expressed by the following formula (32):

[0188] Formula (32);

[0189] in, It is the emergency attenuation coefficient, and ΔTDS is the change in TDS (i.e., the change in total dissolved solids in the makeup water), used to quantify the degree of pollution. This is the initial water distribution volume for replenishment well i. This is the emergency water allocation for replenishment well i.

[0190] The preset maximum allowable water distribution volume can be expressed as Q. safe The preset total dissolved solids safety threshold can be expressed as threshold. TDS If the total dissolved solids in the replenishment water of any of the aforementioned replenishment wells exceed a preset total dissolved solids safety threshold, and the change in the total dissolved solids in the replenishment water is greater than or equal to the change threshold, then... and Q safe The smaller value between the two values ​​is determined as the candidate water distribution volume for replenishment well i.

[0191] If the change amount is less than the change amount threshold, the smaller value between the initial water distribution amount of each replenishment well and the preset maximum allowable water distribution amount is determined as the candidate water distribution amount of the corresponding replenishment well.

[0192] Wherein, if the total dissolved solids in the replenishment water of any of the replenishment wells exceed a preset total dissolved solids safety threshold, and the change in the total dissolved solids in the replenishment water is less than the change threshold, then... and Q safe The smaller value between the two values ​​is determined as the candidate water distribution volume for replenishment well i.

[0193] Step S13092: If the total dissolved solids in the replenishment water of any of the replenishment wells are less than or equal to the preset total dissolved solids safety threshold, the initial water distribution volume of the corresponding replenishment well is determined as the candidate water distribution volume of the corresponding replenishment well.

[0194] Where the total dissolved solids in the replenishment water of any of the replenishment wells are less than or equal to the preset total dissolved solids safety threshold, the initial water distribution volume of the replenishment well can be directly determined as the candidate water distribution volume.

[0195] Step S13093: Rebalance the total amount of candidate water distribution for all replenishment wells to obtain the target water distribution for each replenishment well, so that the sum of the target water distribution for all replenishment wells is equal to the total predicted water volume.

[0196] Among them, the candidate water allocation is calculated separately based on the safety constraints of a single well. This may result in the sum of the candidate water allocation of all replenishment wells not being equal to the total water volume predicted by the model. In this case, total rebalancing is required to correct the deviation, as shown in the following formula (33). Finally, the sum of the target water allocation of all replenishment wells is equal to the total water volume predicted.

[0197] Formula (33);

[0198] in, N represents the target water distribution volume of replenishment well i, and N represents the number of replenishment wells.

[0199] The dynamic water allocation prediction model ensures the safety and feasibility of water allocation through multi-constraint optimization and dynamically responds to system changes through a real-time adjustment mechanism. Specifically, it comprises four core components: water quality safety constraints, water balance constraints, dynamic weight updates, and an emergency response mechanism. When the system detects that the total dissolved solids (TDS) concentration exceeds the safety threshold, it automatically triggers a protection mechanism to restrict water allocation in that area. Simultaneously, the water balance constraint ensures that the total allocated water volume accurately matches the predicted demand, avoiding resource waste. The dynamic weight update continuously optimizes the allocation strategy based on historical recharge efficiency, adjusting the weights of each well location using a gradient descent method. The emergency response mechanism provides a rapid response to sudden changes in water quality. When a sharp change in TDS concentration is detected, the allocated water volume is immediately adjusted exponentially.

[0200] In the above embodiments, an emergency water allocation adjustment mechanism is established with the TDS safety threshold as the core constraint, combined with the change threshold and emergency attenuation coefficient, to avoid the impact risk of water quality exceeding the TDS standard on the groundwater environment. Through two-step optimization of candidate water allocation and total rebalancing, it is ensured that the water allocation of a single well meets the safety constraints, and the sum of the water allocation of all wells is consistent with the total water volume prediction value, taking into account both the safety of a single well and the overall balance. It dynamically adapts to water quality fluctuations and environmental constraints, solves the problems of lack of flexibility and easy ecological risks in traditional water allocation schemes, and improves the safety and reliability of the replenishment process.

[0201] To enhance the model's adaptability to different hydrological features and capture the global dependencies between features, this model designs a feature fusion scheme that combines an adaptive weighting mechanism and a self-attention mechanism. First, the feature extraction capability of the convolutional kernel is optimized through adaptive weight calculation. Then, the similarity between features is calculated using a self-attention mechanism, and dynamic feature fusion is achieved based on attention weights. Finally, water allocation prediction is generated through a fully connected network.

[0202] The decision-making logic of this model is based on the principles of multi-source data fusion and multi-objective optimization, achieving a full-process decision-making process from basic prediction to refined allocation through hierarchical computation. First, basic predicted values ​​are generated based on current hydrological parameters (water level, resistivity) and historical time-series data; then, differentiated allocations are performed according to seasonal characteristics and water quality conditions; finally, multi-constraint optimization ensures the feasibility and safety of the allocation scheme. The specific calculation process is as follows:

[0203] The core algorithm of this model is to construct a water volume prediction framework that can simultaneously capture long-term temporal dependencies and dynamic feature interactions by integrating the dilated causal convolutional structure, gating mechanism, and self-attention mechanism of Temporal Convolutional Networks (TCN). The dilated convolution of TCN ensures an exponentially expanded receptive field, effectively solving the gradient vanishing problem of traditional recurrent neural networks; the gating mechanism controls the information flow through the sigmoid activation function, enhancing the model's ability to filter key features; and the self-attention mechanism dynamically calculates the importance weights of different time steps and feature dimensions, achieving accurate feature fusion. Through multi-level feature abstraction and weighted fusion, this algorithm can extract stable prediction patterns from complex hydrological time-series data, achieving accurate prediction of water replenishment volume.

[0204] In the above embodiments, the gated TCN prediction model, combined with a multi-constraint optimization algorithm, achieves accurate prediction and scientific allocation of recharge water volume. Compared with the traditional average allocation method, it significantly improves recharge efficiency and reduces the risk of groundwater pollution, especially in water-sensitive areas, effectively preventing local water quality deterioration. This optimizes water resource allocation efficiency and reduces environmental risks.

[0205] In the above embodiments, the core dimensions of groundwater hydrological data (water level, resistivity) and recharge water quality data (TDS, COD) are clearly defined to achieve multi-factor coverage of water allocation decisions and overcome the limitations of traditional allocation that only considers a single factor. Temporal convolutional networks extract long-term causal features, gating mechanisms screen key features, self-attention mechanisms focus on the correlation weights between features, and multi-model fusion fully explores the spatiotemporal coupling relationship between hydrological and water quality data to improve the scientific nature of water allocation decisions. The water allocation weight parameters are dynamically updated based on recharge efficiency to achieve resource optimization by allocating more efficient wells and fewer inefficient wells, avoiding waste caused by fixed weight allocation. By combining hydrological response functions and water quality adjustment functions, the original data is transformed into water allocation adaptability features to support accurate water allocation under multiple constraints and improve the utilization rate of recharge water resources and the adaptability of groundwater environment.

[0206] In some embodiments, the method further includes:

[0207] Step S141: Based on the total dissolved solids, chemical oxygen demand and target water distribution of the replenishment water of each replenishment well, a superimposed heat map of water quality and quantity is generated. The heat map reflects the water quality level and water quantity of each replenishment well through different color gradients.

[0208] Step S142: Based on the physical location and fouling level of each ultrafiltration membrane, a three-dimensional fouling distribution map is generated, with different colors corresponding to different fouling levels;

[0209] Step S143: Using the geographical coordinates of each replenishment well as the base and the target water distribution volume of the corresponding replenishment well as the height, a three-dimensional water volume bar chart is generated. The target water distribution volume of each replenishment well is represented by an independent three-dimensional bar.

[0210] The intelligent collaborative control system of this application has developed a dedicated three-dimensional visualization platform and applied a spatial interpolation algorithm, namely the Kriging interpolation method.

[0211] The geographic coordinates (such as latitude and longitude), real-time water quality data (such as TDS), water distribution data, membrane module location information, membrane fouling level data, etc. of the replenishment well are integrated with the Geographic Information System (GIS) base map.

[0212] For discrete well point data, a spatial interpolation algorithm is used to calculate the continuous data field of the entire region, generating a smooth distribution surface.

[0213] When generating the water quality and quantity overlay heat map, the interpolated TDS data field and water distribution data field are overlaid on the electronic map with different colored transparency layers to form an overlay heat map, which intuitively shows where the water quality is poor and where the water volume is large.

[0214] When generating the 3D pollution distribution map, the physical location of the ultrafiltration membrane module is modeled in 3D space, and the pollution level F_level of each module is mapped to color depth or height to form a 3D pollution distribution map, which makes the most polluted area immediately apparent.

[0215] When generating the 3D water volume bar chart, a 3D column is drawn with the target water distribution volume as the height at the geographical location of each replenishment well, so as to intuitively compare the distribution of each well.

[0216] When generating dynamic trend animations, the above visualization results are bound to the timeline to generate GIF (Graphics Interchange Format) animations, which show the dynamic changes in water quality, pollution, and water quantity.

[0217] In the above embodiments, the 3D visualization platform transforms complex data into intuitive spatial distribution heatmaps and 3D dynamic models, significantly improving managers' efficiency in understanding system status, shortening the time for locating abnormal events, increasing the proportion of data-driven decision-making, and significantly reducing operational reliance on personnel experience. This comprehensively enhances system management efficiency and the level of intelligent decision-making.

[0218] The system generates overlay heat maps of water quality and quantity, 3D pollution distribution maps, and 3D water quantity bar charts, transforming abstract water quality data, pollution status, and water distribution results into intuitive visualizations, reducing the difficulty of manual intervention. These visualizations facilitate real-time monitoring of system operation status (membrane fouling distribution, water distribution in each well) by maintenance personnel, enabling them to quickly locate problems and take measures, thus improving the convenience and efficiency of system operation and maintenance. Furthermore, the system provides intuitive guidance for the implementation of water distribution plans, ensuring the traceability and controllability of the replenishment process, and enhancing the overall system operation and maintenance management level.

[0219] The workflow of the intelligent collaborative control system of this application includes the following steps:

[0220] Step S21: The multi-source data acquisition and transmission system collects real-time data from the entire process through a sensor network deployed at each process node. This includes coal mine effluent water quality parameters (pH, total dissolved solids (TDS), turbidity, suspended solids (SS), conductivity (EC), and oil content), membrane system operating parameters (transmembrane pressure difference, product water turbidity, and microbial diversity index), groundwater dynamic data (water level and resistivity), and recharge process data (flow rate, TDS, and COD). This data is transmitted to the central database in real-time via an Industrial Internet of Things (IIoT) protocol, with a collection frequency of once per minute to ensure data timeliness and completeness.

[0221] Step S22: Data preprocessing and quality assurance. The data preprocessing module cleans, corrects, and normalizes the raw data.

[0222] Data cleaning includes identifying and removing outliers using the 3σ criterion and filling in missing data using linear interpolation; data standardization includes applying Min-Max normalization to transform each parameter to the [0,1] interval to eliminate the influence of units; time series alignment includes synchronizing multi-source data based on timestamps to form a standardized time series dataset; quality verification includes real-time calculation of the Data Quality Index (DQI), and triggering a data re-sampling mechanism when DQI < 0.8.

[0223] Step S23: The parallel intelligent analysis and decision-making system executes the analysis tasks of the three models in parallel;

[0224] A sedimentation control model based on attention-enhanced bidirectional LSTM is used as input. Water quality time series data (6-dimensional feature vector) from the past 24 hours is input. The bidirectional LSTM model predicts the optimal influent flow rate, and the attention mechanism identifies key time steps. Influent valve adjustment commands (4 to 20mA control signals) are generated. The opening degree of the sedimentation tank influent valve is adjusted through a PLC (Programmable Logic Controller).

[0225] The membrane fouling early diagnosis model based on variational autoencoder (VAE) takes real-time membrane system sensor data (3D feature vector) as input; the VAE model calculates reconstruction error and potential spatial distribution deviation; outputs fouling status indicator (normal / warning / alarm) and fouling degree score (0-1); triggers cleaning alarm of corresponding level and recommends optimized cleaning scheme.

[0226] Based on a water dynamic allocation prediction model using gated TCN and self-attention, the model takes groundwater hydrological data (water level, resistivity) and recharge water quality data (TDS, COD) as input. The gated TCN model predicts recharge demand, and the self-attention mechanism integrates multi-source features. The model outputs a differentiated allocation scheme for each recharge well and adjusts the water allocation of each well through an intelligent water distribution valve.

[0227] In step S24, the visualization and monitoring early warning module synchronously sends all data, alarm events, and control commands to the 3D visualization engine; it uses the Kriging interpolation algorithm to generate a spatial distribution map of water quality parameters; and it uses WebGL technology to achieve dynamic rendering of the 3D scene (60fps) and displays it in multiple views.

[0228] Water quality heat maps are used to show the spatial distribution of parameters such as TDS and COD;

[0229] Three-dimensional fouling maps are used to show the distribution of membrane fouling levels across the membrane module;

[0230] The water volume bar chart is used to show a three-dimensional comparison of the water volume distribution in each replenishment well;

[0231] Trend curves are used to display the real-time changing trends of key parameters;

[0232] Warning prompts are used to automatically pop up a warning window in abnormal situations, providing handling suggestions.

[0233] Step S25: The closed-loop feedback and continuous optimization system form a complete closed-loop feedback mechanism;

[0234] When monitoring performance, the system calculates indicators such as prediction accuracy and response time of each model in real time; when optimizing parameters, it automatically adjusts model parameters and control strategies based on operational data; when accumulating knowledge, it records expert processing experience and enriches the decision rule base; and when it iterates and upgrades, it automatically retrains the model every month to adapt to system changes.

[0235] This application fundamentally changes the control paradigm of groundwater treatment and replenishment in coal mines, upgrading it from independent, passive, and simple rule-based control to collaborative, proactive, and AI-based intelligent regulation. Its core differences lie in three main aspects:

[0236] At the system architecture level: a multi-model parallel and collaborative intelligent control architecture. Through a data bus and decision center, the sedimentation control model, membrane fouling diagnosis model, and water allocation prediction model are analyzed in parallel and make collaborative decisions, realizing closed-loop optimization of the entire process and solving the control island problem.

[0237] At the core algorithm level: the general model has been improved to address the specific characteristics of water treatment processes.

[0238] In sedimentation control, attention-enhanced bidirectional LSTM is used to solve the problem that traditional models cannot dynamically capture key time-series changes in water quality.

[0239] In membrane fouling diagnosis, variational autoencoders (VAEs) are used for unsupervised learning, enabling early and quantitative diagnosis based on reconstruction errors, overcoming the limitations of relying on fault samples and fixed thresholds.

[0240] In water quality prediction and allocation, a combination of gated TCN and self-attention mechanism is adopted to overcome the shortcomings of traditional time series models, such as limited receptive field and difficulty in capturing long-term dependencies and feature interactions.

[0241] At the decision-making and interaction level: it achieves deep integration of dynamic intelligent decision-making and 3D visualization. In terms of decision-making logic, it adopts dynamic thresholds and trend prediction early warning based on historical data statistics, transforming passive response into proactive intervention; in terms of the interactive interface, it transforms the model output into intuitive water quality / quantity overlaid heat maps and 3D pollution distribution maps through spatial interpolation, providing global and intuitive decision support, far exceeding the traditional two-dimensional chart and numerical list methods.

[0242] In one embodiment, this solution implements intelligent flow control for the settling process. The suspended solids (SS) concentration at the outlet of a coal mine showed a fluctuating upward trend, gradually increasing from an initial 80 mg / L to 180 mg / L, while the safety threshold in this scenario is 150 mg / L, indicating a risk of exceeding the standard.

[0243] The system inputs 24-hour water quality time-series data (including pH, TDS, turbidity, SS, EC, and oil content) into a trained attention-enhanced bidirectional LSTM model. Based on the sequence pattern, the model predicts that the influent flow rate needs to be reduced from 500 m³ / h to 350 m³ / h within the next two hours to ensure sedimentation. The system immediately executes the flow regulation command, adjusting the influent valve opening via the PLC controller.

[0244] In this control operation, the system completed flow regulation 1.5 hours in advance. Subsequent monitoring showed that the suspended solids (SS) concentration in the sedimentation tank effluent remained stable within the range of 90-110 mg / L, with no exceedances exceeding 150 mg / L. In contrast, traditional threshold control methods require waiting for the measured SS value to exceed the standard before taking action, leading to a temporary deterioration in effluent quality (SS peak reaching 200 mg / L) and a longer recovery time. This case demonstrates the advantages of the LSTM model in proactive prediction and fine-grained control, significantly improving the stability and shock resistance of the sedimentation process.

[0245] In one embodiment, this solution performs early diagnosis and cleaning optimization of membrane fouling. During the operation of the ultrafiltration membrane system, the transmembrane pressure difference (TMP) shows a slight upward trend (from 0.05MPa to 0.06MPa), but does not exceed the traditional alarm threshold of 0.08MPa, and the turbidity of the permeate does not change significantly.

[0246] The system inputs real-time sensor data (TMP, permeate turbidity, and microbial diversity index) into the variational autoencoder (VAE) model. The model calculates a reconstruction error of 0.39 (the preset normal threshold is 0.1) and a potential spatial deviation of 0.45, which is considered a slight contamination. The system immediately triggers an alarm and recommends a chemical cleaning solution, estimated to take 2 hours, with the cleaning agent dosage being 80% of the standard dose.

[0247] In this cleaning process, management promptly executed the cleaning procedure, and the membrane performance was fully restored after cleaning (TMP dropped back to 0.05 MPa). Traditional methods require waiting for TMP to rise above 0.08 MPa before initiating cleaning, which often leads to more severe contamination and necessitates a combined physicochemical cleaning approach (taking 6 to 8 hours and increasing cleaning agent usage by 50%). This case demonstrates the effectiveness of the VAE model in early diagnosis and quantitative assessment, preventing irreversible degradation of membrane performance and reducing maintenance costs and downtime.

[0248] In one embodiment, this solution performs intelligent water allocation during the dry season. During the dry season, the groundwater level is generally low (average water level 48.5m, lower than the baseline water level of 50m), and the TDS monitoring value near a certain replenishment well is high (monitoring value 650mg / L, baseline value 400mg / L).

[0249] The system inputs groundwater hydrological data (water level, resistivity) and recharge water quality data (TDS, COD) into a gated TCN and a self-attention model. The model predicts a total water demand of 800 m³ / h and, based on a water quality constraint strategy, automatically assigns a weight reduction factor of 0.7 to well locations with high TDS, while other well locations are assigned normal weights.

[0250] In this allocation, the system generated a differentiated allocation scheme, allocating 56 m³ / h (7% of the total) to wells with high TDS, and allocating between 90-110 m³ / h to other wells. After execution, the overall recharge demand was met, and no further pollution occurred in the high TDS area (TDS stabilized at 650-670 mg / L). A traditional average allocation scheme would have allocated 100 m³ / h to the well, potentially exacerbating local pollution. This case demonstrates the model's capabilities in multi-constraint optimization and refined allocation, improving recharge efficiency and reducing environmental risks.

[0251] Based on the foregoing embodiments, this application further provides an intelligent collaborative control system for coal mine groundwater treatment and replenishment. The system includes various modules and sub-modules, and each unit of each sub-module can be implemented by a processor in an electronic device; of course, it can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0252] Figure 5 This is a schematic diagram of the composition structure of an intelligent collaborative control system for coal mine groundwater treatment and replenishment provided in an embodiment of this application, as shown below. Figure 5 As shown, the system 500 includes:

[0253] The data perception layer 51, the intelligent analysis layer 52, and the decision execution layer 53, wherein:

[0254] The data perception layer 51 is used to acquire the time series data of the influent water quality of the raw mine water generated by coal mining. The intelligent analysis layer 52 is used to predict the optimal influent flow rate based on the influent water quality time series data using a bidirectional long short-term memory network. The decision execution layer 53 is used to adjust the influent flow rate of the sedimentation tank to the optimal influent flow rate in order to perform sedimentation pretreatment on the raw mine water.

[0255] The data sensing layer 51 is used to obtain multidimensional membrane operating parameters of the ultrafiltration membrane after sedimentation pretreatment. The intelligent analysis layer 52 is used to determine the degree of fouling of the ultrafiltration membrane based on the multidimensional membrane operating parameters using a trained variational autoencoder, and dynamically adjust the cleaning strategy of the ultrafiltration membrane based on the degree of fouling. The decision execution layer 53 is used to execute the cleaning operation of the ultrafiltration membrane based on the cleaning strategy.

[0256] The data sensing layer 51 is used to acquire groundwater hydrological data of the target replenishment area and water quality data of the replenished water that meets the standards of the ultrafiltration membrane after cleaning. The intelligent analysis layer 52 is used to determine the allocation scheme of each replenishment well in the target replenishment area based on the groundwater hydrological data and the water quality data of the water dynamic allocation prediction model that integrates temporal convolutional networks, gating mechanisms and self-attention mechanisms. The decision execution layer 53 is used to replenish the compliant water to the target groundwater layer through the corresponding replenishment wells according to the allocation scheme of each replenishment well.

[0257] In some possible embodiments, the influent water quality time-series data includes pH value, total dissolved solids, turbidity, suspended solids concentration, conductivity, and oil content; the intelligent analysis layer 52 is used to: extract forward temporal features of the influent water quality time-series data using the forward LSTM encoding layer of the bidirectional long short-term memory network, extract backward temporal features of the influent water quality time-series data using the backward LSTM encoding layer of the bidirectional long short-term memory network, concatenate the forward and backward temporal features according to the time step dimension to obtain bidirectional concatenated temporal features, wherein the bidirectional concatenated temporal features include feature vectors corresponding to multiple time steps; and utilize the bidirectional long short-term memory network... The attention score vector, attention weight matrix, attention bias, and the bidirectional splicing temporal features of the attention mechanism layer are used to determine the attention weights of each time step in the bidirectional splicing temporal features. Based on the attention weights of each time step, the feature vectors of multiple time steps in the bidirectional splicing temporal features are weighted and fused to obtain a weighted feature vector. Using the fully connected weight matrix, fully connected bias, and the weighted feature vector of the fully connected prediction layer of the bidirectional long short-term memory network, the initial influent flow rate is determined through linear transformation and activation function calculation. Based on the initial influent flow rate, the suspended solids concentration, the preset baseline flow rate, and the preset suspended solids safety threshold, the optimal influent flow rate is determined.

[0258] In some possible embodiments, the intelligent analysis layer 52 is further configured to: determine the smaller of the product of the initial influent flow rate, the preset reference flow rate, and the preset proportional coefficient as the optimal influent flow rate when the suspended solids concentration is greater than the preset suspended solids safety threshold, wherein the preset proportional coefficient is less than 1; determine the larger of the initial influent flow rate and the preset reference flow rate as the optimal influent flow rate when the suspended solids concentration is less than or equal to the preset suspended solids safety threshold and the initial influent flow rate is greater than or equal to the preset reference flow rate; and determine the initial influent flow rate as the optimal influent flow rate when the suspended solids concentration is less than or equal to the preset suspended solids safety threshold and the initial influent flow rate is less than the preset reference flow rate.

[0259] In some possible embodiments, the multidimensional membrane operating parameters include transmembrane pressure difference, permeate turbidity, and microbial diversity index. The intelligent analysis layer 52 is used to: extract depth features from the multidimensional membrane operating parameters using a three-layer fully connected network of the probabilistic encoder of the trained variational autoencoder, and output the mean and variance of the posterior Gaussian distribution of the latent space; obtain latent variables based on the mean and variance through reparameterization sampling using the reparameterization sampling module of the variational autoencoder; and use a three-layer fully connected network symmetrical to the three-layer fully connected network of the probabilistic decoder of the variational autoencoder to analyze the... The latent variables are reconstructed to obtain reconstructed data of the multidimensional membrane operating parameters. Using the output layer of the variational autoencoder, the reconstruction probability of the multidimensional membrane operating parameters is determined based on the reconstructed data and the multidimensional membrane operating parameters. The relative entropy between the approximate posterior distribution and the prior distribution of the latent variables is determined, and this relative entropy is defined as the latent spatial deviation. Based on the reconstruction probability, the latent spatial deviation, a preset weighting coefficient, and a preset maximum latent spatial deviation, an anomaly score is determined. Based on the anomaly score, a preset normal period mean, and a preset normal period standard deviation, the fouling degree of the ultrafiltration membrane is determined.

[0260] In some possible embodiments, the intelligent analysis layer 52 is further configured to determine a normal threshold and a warning threshold based on the preset normal period mean and the preset normal period standard deviation; determine that the ultrafiltration membrane's fouling level is normal operation when the abnormal score is less than the normal threshold; determine that the ultrafiltration membrane's fouling level is a mild fouling warning when the abnormal score is greater than or equal to the normal threshold and less than the warning threshold; and determine that the ultrafiltration membrane's fouling level is a severe fouling warning when the abnormal score is greater than or equal to the warning threshold.

[0261] In some possible embodiments, the intelligent analysis layer 52 is further configured to: input historical multidimensional membrane operating parameters from the training dataset into the initial variational autoencoder to be trained to obtain corresponding historical reconstruction data; determine the reconstruction loss during training based on the historical reconstruction data and the historical multidimensional membrane operating parameters; determine the relative entropy loss during training based on the mean and variance of the output of the initial variational autoencoder, and a preset prior mean and prior variance; determine the temporal consistency loss during training based on the temporal sequence corresponding to the latent variables output by the initial variational autoencoder; determine the value of the multi-objective loss function by weighted fusion using preset weight coefficients based on the reconstruction loss, the relative entropy loss, and the temporal consistency loss; and iteratively update the initial variational autoencoder using gradient descent based on the value of the multi-objective loss function until the value of the multi-objective loss function converges, thereby obtaining the trained variational autoencoder.

[0262] In some possible embodiments, the groundwater hydrological data includes groundwater level and resistivity, and the recharge water quality data includes total dissolved solids and chemical oxygen demand (COD) of the recharge water from each recharge well; the intelligent analysis layer 52 is used to: extract long-term causal features from the groundwater hydrological data by sampling time-series data at intervals, based on the dilation factor, kernel size, and kernel weight parameters of each dilated causal convolutional layer in a time-series convolutional network; perform feature transformation on the long-term causal features based on the feature transformation convolution kernel weight matrix to obtain transformed features, and perform feature transformation based on gated convolution kernels. A weight matrix and the long-term causal features are used to generate a gating signal. Based on the transformed features and the gating signal, key features strongly correlated with membrane fouling are selected through a gating mechanism. An adaptive importance weight is determined for each convolutional kernel based on the adaptive weight matrix, weight projection vector, and weight parameters of the convolutional kernels. High-importance convolutional kernels are selected based on their adaptive importance weights, and feature extraction is performed on the key features using these high-importance kernels to obtain high-quality feature vectors. Finally, any two high-quality feature vectors are mapped to a lookup weight matrix and a key weight matrix. The query vector and key vector are used; based on the inner product of the query vector and the key vector and the square root of the dimension of the key vector, the cross-attention weights between temporal features are obtained; based on the value weight matrix, the high-quality feature vector at each time step is converted into a value vector; based on the cross-attention weights, the value vectors at multiple time steps are weighted and fused to obtain the attention-weighted fusion feature corresponding to each replenishment well; based on the attention-weighted fusion feature corresponding to each replenishment well, the hidden layer weight matrix and the hidden layer bias vector, the output layer weight matrix and the output layer bias vector, the predicted total water volume of the target replenishment area is determined; based on The hydrological response function value is determined based on the groundwater hydrological data and hydrological response function. The water quality adjustment function value is determined based on the recharge water quality data and water quality adjustment function. Based on the total predicted water volume, the water allocation weight parameter for each recharge well, the hydrological response function value, and the water quality adjustment function value, the total predicted water volume is allocated to each recharge well to obtain the initial water allocation for each recharge well. The water allocation weight parameter for each recharge well is dynamically updated based on the recharge efficiency of the corresponding recharge well. Based on a multi-constraint optimization strategy, the initial water allocation for each recharge well is optimized to obtain the target allocation scheme for the corresponding recharge well.

[0263] In some possible embodiments, the intelligent analysis layer 52 is configured to: determine the change in the total dissolved solids (TDS) of the replenishment water of any of the replenishment wells when the TDS of the replenishment water exceeds a preset TDS safety threshold; adjust the initial water distribution of the corresponding replenishment well based on the change in TDS and the emergency attenuation coefficient when the change is greater than or equal to the change threshold, thereby obtaining the emergency water distribution of the corresponding replenishment well; and determine the smaller value between the emergency water distribution of the corresponding replenishment well and the preset maximum allowable water distribution as the reserve water distribution for the corresponding replenishment well. Select the water allocation amount; if the change amount is less than the change amount threshold, determine the smaller value between the initial water allocation amount of each replenishment well and the preset maximum allowable water allocation amount as the candidate water allocation amount of the corresponding replenishment well; if the total dissolved solids in the replenishment water of any replenishment well are less than or equal to the preset total dissolved solids safety threshold, determine the initial water allocation amount of the corresponding replenishment well as the candidate water allocation amount of the corresponding replenishment well; rebalance the total amount of candidate water allocation amounts of all replenishment wells to obtain the target water allocation amount of each replenishment well, so that the sum of the target water allocation amounts of all replenishment wells is equal to the total water volume prediction value.

[0264] In some possible embodiments, the system further includes a visualization layer for: generating a water quality and quantity overlay heat map based on the total dissolved solids, chemical oxygen demand (COD), and target water distribution volume of the replenishment water in each replenishment well; the heat map reflecting the water quality level and water volume of each replenishment well through different color gradients; generating a three-dimensional pollution distribution map based on the physical location and pollution level of each ultrafiltration membrane, with different pollution levels corresponding to different colors; and generating a three-dimensional water volume bar chart based on the geographical coordinates of each replenishment well and the target water distribution volume of the corresponding replenishment well as the height, with the target water distribution volume of each replenishment well represented by an independent three-dimensional bar.

[0265] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0266] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0267] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0268] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of the embodiments of this application according to actual needs. In addition, each functional unit in the embodiments of this application may be fully integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in the form of hardware plus software functional units.

[0269] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0270] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined to obtain new method embodiments without conflict. The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined to obtain new method embodiments or device embodiments without conflict.

[0271] The above description is merely an embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart collaborative control method for coal mine groundwater treatment and recharge, the method comprising: The time-series data of the influent water quality of the raw mine water generated by coal mining is obtained. Based on the time-series data of the influent water quality, a bidirectional long short-term memory network is used to predict the optimal influent flow rate. The influent flow rate of the sedimentation tank is adjusted to the optimal influent flow rate to perform sedimentation pretreatment on the raw mine water. After sedimentation pretreatment, multidimensional membrane operating parameters of the ultrafiltration membrane are obtained. The degree of fouling of the ultrafiltration membrane is determined based on the multidimensional membrane operating parameters using a trained variational autoencoder. The cleaning strategy of the ultrafiltration membrane is dynamically adjusted based on the degree of fouling. The cleaning operation of the ultrafiltration membrane is performed based on the cleaning strategy. The multidimensional membrane operating parameters include transmembrane pressure difference, permeate turbidity, and microbial diversity index. After cleaning, groundwater hydrological data of the target replenishment area and water quality data of the replenished water from the ultrafiltration membrane are acquired. A dynamic water allocation prediction model, incorporating temporal convolutional networks, gating mechanisms, and self-attention mechanisms, is used to determine the allocation scheme for each replenishment well within the target replenishment area based on the groundwater hydrological data and the replenished water quality data. The compliant effluent is then replenished to the target groundwater layer through the corresponding replenishment wells according to the allocation scheme. The process of determining the allocation scheme for each replenishment well within the target replenishment area using the dynamic water allocation prediction model, incorporating temporal convolutional networks, gating mechanisms, and self-attention mechanisms, includes: Based on the dilation factor, kernel size, and kernel weight parameters of each dilated causal convolutional layer in a temporal convolutional network, long-term temporal causal features are extracted from the groundwater hydrological data by sampling temporal data at intervals. Based on the feature transformation convolution kernel weight matrix, the long-term causal features are transformed to obtain the transformed features. Based on the gated convolution kernel weight matrix and the long-term causal features, a gated signal is generated. Based on the transformed features and the gated signal, key features strongly correlated with membrane fouling are screened out through the gated mechanism. Based on the adaptive weight matrix, weight projection vector, and weight parameters of the convolution kernel, the adaptive importance weight of each convolution kernel is determined; based on the adaptive importance weight of each convolution kernel, the convolution kernels are filtered to obtain high-importance convolution kernels; based on the high-importance convolution kernels, the key features are extracted to obtain high-quality feature vectors. Based on the query weight matrix and the key weight matrix, any two high-quality feature vectors are mapped to query vector and key vector respectively; based on the inner product of the query vector and the key vector and the square root of the dimension of the key vector, the cross-attention weights between temporal features are obtained. Based on the value weight matrix, the high-quality feature vector at each time step is converted into a value vector. Based on the cross attention weight, the value vectors at multiple time steps are weighted and fused to obtain the attention-weighted fusion feature corresponding to each backfill well. Based on the attention-weighted fusion features, hidden layer weight matrix and hidden layer bias vector, output layer weight matrix and output layer bias vector corresponding to each replenishment well, the total water volume prediction value of the target replenishment area is determined; Based on the groundwater hydrological data and hydrological response function, the hydrological response function value is determined; based on the recharge water quality data and water quality adjustment function, the water quality adjustment function value is determined. Based on the total water volume prediction, the water allocation weight parameter of each replenishment well, the hydrological response function value, and the water quality adjustment function value, the total water volume prediction is allocated to each replenishment well to obtain the initial water allocation of the corresponding replenishment well. The water allocation weight parameter of each replenishment well is dynamically updated based on the replenishment efficiency of the corresponding replenishment well. Based on a multi-constraint optimization strategy, the initial water distribution of each replenishment well is optimized to obtain the target allocation scheme for the corresponding replenishment well.

2. The method according to claim 1, characterized in that, The influent water quality time-series data includes pH value, total dissolved solids, turbidity, suspended solids concentration, conductivity, and oil content; the method of predicting the optimal influent flow rate based on the influent water quality time-series data using a bidirectional long short-term memory network includes: The forward temporal features of the influent water quality time series data are extracted using the forward LSTM encoding layer of the bidirectional long short-term memory network, and the backward temporal features of the influent water quality time series data are extracted using the backward LSTM encoding layer of the bidirectional long short-term memory network. The forward and backward temporal features are concatenated along the time step dimension to obtain bidirectional concatenated temporal features, which include feature vectors corresponding to multiple time steps. Using the attention score vector, attention weight matrix, attention bias, and the bidirectional splicing temporal features of the attention mechanism layer, the attention weight of each time step in the bidirectional splicing temporal features is determined; based on the attention weight of each time step, the feature vectors of multiple time steps in the bidirectional splicing temporal features are weighted and fused to obtain a weighted feature vector. Using the fully connected weight matrix, fully connected bias, and the weighted eigenvector of the output layer, the initial influent flow rate is determined through linear transformation and activation function calculation. The optimal influent flow rate is determined based on the initial influent flow rate, the suspended solids concentration, the preset baseline flow rate, and the preset suspended solids safety threshold.

3. The method according to claim 2, characterized in that, The process of determining the optimal influent flow rate based on the initial influent flow rate, the suspended solids concentration, the preset baseline flow rate, and the preset suspended solids safety threshold includes: When the suspended solids concentration is greater than the preset suspended solids safety threshold, the smaller value among the initial influent flow rate, the preset baseline flow rate, and the preset proportional coefficient is determined as the optimal influent flow rate, where the preset proportional coefficient is less than 1. When the suspended solids concentration is less than or equal to the preset suspended solids safety threshold and the initial influent flow rate is greater than or equal to the preset reference flow rate, the larger of the initial influent flow rate and the preset reference flow rate is determined as the optimal influent flow rate. When the suspended solids concentration is less than or equal to the preset suspended solids safety threshold and the initial influent flow rate is less than the preset reference flow rate, the initial influent flow rate is determined as the optimal influent flow rate.

4. The method according to claim 1, characterized in that, The determination of the fouling level of the ultrafiltration membrane based on the multidimensional membrane operating parameters using the trained variational autoencoder includes: Using a three-layer fully connected network of a probabilistic encoder that has been trained variational autoencoder, depth features are extracted from the multidimensional membrane operating parameters, and the mean and variance of the posterior Gaussian distribution of the latent space are output. Using the reparameterization sampling module of the variational autoencoder, latent variables are obtained through reparameterization sampling based on the mean and the variance; using a three-layer fully connected network in the probabilistic decoder of the variational autoencoder that is symmetrical to the three-layer fully connected network of the probabilistic encoder, the latent variables are reconstructed to obtain the reconstructed data of the multidimensional membrane operating parameters. Using the output layer of the variational autoencoder, based on the reconstructed data and the multidimensional membrane operating parameters, the reconstruction probability of the multidimensional membrane operating parameters is determined; the relative entropy between the approximate posterior distribution of the latent variable and the prior distribution of the latent variable is determined, and the relative entropy is determined as the latent space deviation. Anomaly scores are determined based on the reconstruction probability, the latent space deviation, the preset weight coefficient, and the preset maximum latent space deviation. The degree of fouling of the ultrafiltration membrane is determined based on the abnormal score, the preset normal period mean, and the preset normal period standard deviation.

5. The method according to claim 4, characterized in that, The determination of the fouling degree of the ultrafiltration membrane based on the anomaly score, the preset normal period mean, and the preset normal period standard deviation includes: Based on the preset normal period mean and the preset normal period standard deviation, the normal threshold and warning threshold are determined; If the abnormal score is less than the normal threshold, the fouling level of the ultrafiltration membrane is determined to be normal operation. If the abnormal score is greater than or equal to the normal threshold and less than the warning threshold, the fouling level of the ultrafiltration membrane is determined to be a mild fouling warning. If the abnormal score is greater than or equal to the warning threshold, the fouling level of the ultrafiltration membrane is determined to be a severe fouling warning.

6. The method according to claim 4, characterized in that, The method further includes: The historical multidimensional membrane running parameters in the training dataset are input into the initial variational autoencoder to be trained to obtain the corresponding historical reconstructed data. Based on the historical reconstruction data and the historical multidimensional membrane operating parameters, the reconstruction loss during the training process is determined; Based on the mean and variance of the initial variational autoencoder output, and the preset prior mean and prior variance, the relative entropy loss during the training process is determined. Based on the temporal sequence corresponding to the latent variables output by the initial variational autoencoder, the temporal consistency loss during the training process is determined. Based on the reconstruction loss, the relative entropy loss, and the temporal consistency loss, the value of the multi-objective loss function is determined by weighted fusion using preset weight coefficients; Based on the value of the multi-objective loss function, the initial variational autoencoder is iteratively updated using the gradient descent method until the value of the multi-objective loss function converges, thus obtaining the trained variational autoencoder.

7. The method according to claim 1, characterized in that, The groundwater hydrological data includes groundwater level and resistivity, and the replenishment water quality data includes total dissolved solids and chemical oxygen demand of the replenishment water in each replenishment well.

8. The method according to claim 7, characterized in that, The multi-constraint optimization strategy optimizes the initial water distribution of each replenishment well to obtain the target allocation scheme for the corresponding replenishment well, including: If the total dissolved solids in the replenishment water of any of the aforementioned replenishment wells exceed a preset total dissolved solids safety threshold... Determine the change in total dissolved solids in the replenishment water of the corresponding replenishment well; if the change is greater than or equal to the change threshold, adjust the initial water distribution of the corresponding replenishment well based on the change in total dissolved solids in the replenishment water of each replenishment well and the emergency attenuation coefficient to obtain the emergency water distribution of the corresponding replenishment well; determine the smaller value between the emergency water distribution of the corresponding replenishment well and the preset maximum allowable water distribution as the candidate water distribution of the corresponding replenishment well; If the change amount is less than the change amount threshold, the smaller value between the initial water distribution amount of each replenishment well and the preset maximum allowable water distribution amount is determined as the candidate water distribution amount of the corresponding replenishment well. If the total dissolved solids in the replenishment water of any of the replenishment wells are less than or equal to the preset total dissolved solids safety threshold, the initial water allocation volume of the corresponding replenishment well will be determined as the candidate water allocation volume of the corresponding replenishment well. The total water allocation of all replenishment wells is rebalanced to obtain the target water allocation for each replenishment well, so that the sum of the target water allocation for all replenishment wells is equal to the total predicted water volume.

9. The method according to claim 1, characterized in that, The method further includes: Based on the total dissolved solids, chemical oxygen demand and target water distribution of the replenishment water of each replenishment well, a superimposed heat map of water quality and quantity is generated. The heat map reflects the water quality level and water quantity of each replenishment well through different color gradients. Based on the physical location and degree of fouling of each ultrafiltration membrane, a three-dimensional fouling distribution map is generated, with different colors corresponding to different degrees of fouling; Using the geographical coordinates of each replenishment well as the base and the target water distribution volume of the corresponding replenishment well as the height, a three-dimensional water volume bar chart is generated. The target water distribution volume of each replenishment well is represented by an independent three-dimensional bar.

10. An intelligent collaborative control system for coal mine groundwater treatment and replenishment, the system comprising: The data perception layer, intelligent analysis layer, and decision execution layer are as follows: The data perception layer is used to acquire the time-series data of the influent water quality of the raw mine water generated by coal mining. The intelligent analysis layer is used to predict the optimal influent flow rate based on the influent water quality time-series data using a bidirectional long short-term memory network. The decision execution layer is used to adjust the influent flow rate of the sedimentation tank to the optimal influent flow rate in order to perform sedimentation pretreatment on the raw mine water. The data sensing layer is used to acquire multidimensional membrane operating parameters of the ultrafiltration membrane after sedimentation pretreatment. The intelligent analysis layer is used to determine the fouling degree of the ultrafiltration membrane based on the multidimensional membrane operating parameters using a trained variational autoencoder, and dynamically adjust the cleaning strategy of the ultrafiltration membrane based on the fouling degree. The decision execution layer is used to execute the cleaning operation of the ultrafiltration membrane based on the cleaning strategy. The multidimensional membrane operating parameters include transmembrane pressure difference, permeate turbidity, and microbial diversity index. The data sensing layer is used to acquire groundwater hydrological data of the target replenishment area and replenishment water quality data of the ultrafiltration membrane effluent after cleaning. The intelligent analysis layer is used to determine the allocation scheme of each replenishment well in the target replenishment area based on the groundwater hydrological data and the replenishment water quality data using a water dynamic allocation prediction model that integrates temporal convolutional networks, gating mechanisms, and self-attention mechanisms. The decision execution layer is used to replenish the target groundwater layer with the effluent according to the allocation scheme of each replenishment well through the corresponding replenishment well. The intelligent analysis layer is used for: Based on the dilation factor, kernel size, and kernel weight parameters of each dilated causal convolutional layer in a temporal convolutional network, long-term temporal causal features are extracted from the groundwater hydrological data by sampling temporal data at intervals. Based on the feature transformation convolution kernel weight matrix, the long-term causal features are transformed to obtain the transformed features. Based on the gated convolution kernel weight matrix and the long-term causal features, a gated signal is generated. Based on the transformed features and the gated signal, key features strongly correlated with membrane fouling are screened out through the gated mechanism. Based on the adaptive weight matrix, weight projection vector, and weight parameters of the convolution kernel, the adaptive importance weight of each convolution kernel is determined; based on the adaptive importance weight of each convolution kernel, the convolution kernels are filtered to obtain high-importance convolution kernels; based on the high-importance convolution kernels, the key features are extracted to obtain high-quality feature vectors. Based on the query weight matrix and the key weight matrix, any two high-quality feature vectors are mapped to query vector and key vector respectively; based on the inner product of the query vector and the key vector and the square root of the dimension of the key vector, the cross-attention weights between temporal features are obtained. Based on the value weight matrix, the high-quality feature vector at each time step is converted into a value vector. Based on the cross attention weight, the value vectors at multiple time steps are weighted and fused to obtain the attention-weighted fusion feature corresponding to each backfill well. Based on the attention-weighted fusion features, hidden layer weight matrix and hidden layer bias vector, output layer weight matrix and output layer bias vector corresponding to each replenishment well, the total water volume prediction value of the target replenishment area is determined; Based on the groundwater hydrological data and hydrological response function, the hydrological response function value is determined; based on the recharge water quality data and water quality adjustment function, the water quality adjustment function value is determined. Based on the total water volume prediction, the water allocation weight parameter of each replenishment well, the hydrological response function value, and the water quality adjustment function value, the total water volume prediction is allocated to each replenishment well to obtain the initial water allocation for the corresponding replenishment well. The water allocation weight parameter of each replenishment well is dynamically updated based on the replenishment efficiency of the corresponding replenishment well. Based on a multi-constraint optimization strategy, the initial water allocation for each replenishment well is optimized to obtain the target allocation scheme for the corresponding replenishment well.

Images