Tpt empowerment intelligent dosing treatment method and system based on time sequence large model

Abstract

This invention provides a time-series large-scale model-based TPT-enabled intelligent chemical dosing method and system, relating to the fields of big data analysis and water treatment technology. It collects multi-dimensional operating parameters in real time through a TPT IoT system, forming a multi-source time-series dataset. After preprocessing at the edge, the dataset is transmitted to a cloud analysis platform. The time-series prediction model is used to mine time-related features, predict water quality change trends, and output the optimal combination of dosing parameters. This enables accurate prediction and efficient control of complex water quality fluctuations, improving the precision and efficiency of chemical dosing. Through a closed-loop feedback optimization mechanism, dosing parameters can be dynamically corrected, further enhancing the system's stability and responsiveness. This effectively solves the problems of low data processing efficiency, insufficient prediction accuracy, and lack of dynamic optimization in existing technologies, significantly improving the intelligence and reliability of the water treatment process.

Description

Technical Field

[0001] This invention relates to the fields of big data analysis and water treatment technology, and more specifically, to a TPT-enabled intelligent dosing treatment method and system based on a time-series large model. Background Technology

[0002] In recent years, with the acceleration of industrialization and urbanization, water pollution has become increasingly serious, highlighting the growing importance of water treatment technology. Traditional water treatment dosing methods rely heavily on manual experience or simple rule-based control, which has limited responsiveness to complex water quality changes and cannot meet the demands of modern water treatment. Emerging water treatment models leveraging advanced technologies such as the Internet of Things (IoT), Artificial Intelligence (AI), and big data analytics have gradually become research hotspots. Among these, intelligent dosing control methods based on time-series data analysis have received widespread attention due to their efficiency and accuracy in dynamically changing environments. In recent years, the development of time-series data processing technology has provided crucial technical support for optimizing water treatment processes, particularly demonstrating significant application potential in predicting water quality changes, optimizing dosing parameters, and achieving closed-loop control. However, in practical applications, due to the complexity of multi-source data, the dynamic interactivity of multi-dimensional parameters, and the high requirements for real-time response in water treatment scenarios, existing technologies still have many shortcomings in terms of efficiency, accuracy, and stability.

[0003] Traditional intelligent dosing technologies for water treatment primarily rely on parameter optimization based on single data sources or static models. This approach often ignores the time-series correlation characteristics of water quality changes, resulting in insufficient predictive power for complex water quality fluctuations. Furthermore, most existing dosing parameter optimization methods employ offline analysis, making it difficult to efficiently process and dynamically feedback real-time data. This leads to unstable dosing effects and may even result in resource waste or secondary pollution. The shortcomings of existing technologies in data preprocessing, model prediction capabilities, and closed-loop optimization mechanisms are specifically reflected in the following aspects: First, low data acquisition and processing efficiency hinders the effective integration of multi-source heterogeneous data, making it difficult to form a comprehensive understanding of water quality changes. Second, time-series prediction models lack sufficient depth in mining dynamic change characteristics, resulting in low accuracy and reliability of prediction results. Third, the lack of a closed-loop optimization mechanism prevents dynamic adjustments to the dosing strategy based on post-dosing effect data. These technical bottlenecks severely restrict the practical application of intelligent dosing technologies for water treatment in complex environments. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a TPT-enabled intelligent dosing treatment method and system based on a time-series large model.

[0005] According to one aspect of the present invention, a TPT-enabled intelligent dosing treatment method based on a time-series large model is provided, comprising: The TPT IoT system collects multi-dimensional operating parameters in real time to form a multi-source time-series dataset, which is then preprocessed at the edge and transmitted to the cloud analysis platform through edge-cloud collaboration. The multi-source time-series dataset is input into the time-series prediction model to mine time-related features, predict water quality change trends based on the mined features, and output the optimal combination of dosing parameters. The optimal combination of dosing parameters is converted into control commands and sent to the automatic dosing equipment. Real-time data on the treatment effect after dosing is collected and compared with the expected target to generate parameter deviation correction values. The processing effect data and the deviation correction value are fed back to the time series prediction model for iterative optimization, updating the time-related feature weights and prediction parameters.

[0006] Furthermore, mining the time-related features includes the following steps: Multi-source time-series data are processed in layers by setting sliding windows with different time granularities; By combining multi-head attention mechanism to calculate dynamic correlation weights and time delay features between variables, and using convolutional neural network to extract local time patterns and identify periodic patterns through frequency domain analysis; Ultimately, causal relationship analysis and mutual information calculation are integrated to comprehensively uncover the complex temporal correlation features in time series data.

[0007] Furthermore, the multi-head attention mechanism identifies causal relationships and time delays between variables by mapping time-series data into query-key-value triples and calculating attention scores at different time offsets. Each attention head captures short-term direct, medium-term transmission, and long-term cumulative causal effects, and finally generates a quantified causal correlation matrix through softmax normalization.

[0008] Furthermore, the optimal dosing parameter combination is output by solving the Pareto optimal solution set through a multi-objective optimization algorithm. Combined with the current operating conditions and constraints, the analytic hierarchy process is used to comprehensively evaluate and output a structured optimal dosing parameter combination that includes specific drug types, precise dosage, and detailed dosing frequency.

[0009] Furthermore, the multi-objective optimization algorithm includes phased initialization and iterative solution; The phased initialization includes taking water quality compliance rate, treatment cost, chemical consumption and equipment energy consumption as the main optimization objectives, while setting upper and lower limits for chemical dosage, chemical inventory constraints and equipment treatment capacity limits as constraints. The iterative solution employs a non-dominated sorting method to stratify all candidate solutions according to Pareto dominance.

[0010] Furthermore, the multi-objective optimization algorithm is shown in the following equation: , in, This is the evaluation index for Pareto optimal solutions. Let be the weighting coefficient for the k-th objective, where k includes the water quality compliance rate, treatment cost, reagent consumption, and equipment energy consumption. For the k-th objective function value, Let be the expected value of the k-th objective. For non-dominated ordering decay parameters, To solve for the non-dominated ordering level of x Adjust parameters for crowding distance. For individual i, the crowding distance To constrain the penalty coefficient, The total number of decision variables, For the j-th constraint violation function, Let j be the current value of the j-th decision variable. Let j be the upper bound of the decision variable. Let be the lower bound of the j-th decision variable.

[0011] Furthermore, the optimal combination of dosing parameters is converted into control commands. The dosing type parameters are converted into equipment start-stop commands through a reagent selection mapping table. The dosing dosage parameters are converted into inverter frequency setpoints based on the pump characteristic curve and reagent density. The dosing frequency parameters are converted into timing control commands. Finally, all control parameters are encapsulated into standard data frames and checksums are added according to the target equipment's communication protocol.

[0012] According to another aspect of the present invention, a TPT-enabled intelligent dosing system based on a time-series large model is provided, comprising: The data acquisition and transmission module is used to collect multi-dimensional operating parameters in real time through the TPT IoT system, form a multi-source time-series dataset, perform preprocessing on the edge side, and transmit it to the cloud analysis platform through edge-cloud collaboration. The intelligent prediction and optimization module is used to input multi-source time-series datasets into the time-series prediction model, mine time-related features, predict water quality change trends based on the mined features, and output the optimal combination of dosing parameters. The control execution monitoring module is used to convert the optimal combination of dosing parameters into control commands and send them to the automatic dosing equipment. It also collects the treatment effect data after dosing in real time and compares it with the expected target to generate parameter deviation correction values. The feedback learning optimization module is used to feed back the processing effect data and bias correction values ​​to the time series prediction model for iterative optimization, and update the weights of time-related features and prediction parameters.

[0013] Compared with existing technologies, the TPT-enabled intelligent chemical dosing method and system based on a time-series large model provided by this invention collects multi-dimensional operating parameters in real time through a TPT IoT system, forming a multi-source time-series dataset. After preprocessing at the edge, the dataset is transmitted to a cloud analysis platform. The time-series prediction model is used to mine time-related features, predict water quality change trends, and output the optimal combination of dosing parameters. This enables accurate prediction and efficient control of complex water quality fluctuations, improving the precision and efficiency of chemical dosing. Through a closed-loop feedback optimization mechanism, dosing parameters can be dynamically corrected, further enhancing the system's stability and responsiveness. This effectively solves the problems of low data processing efficiency, insufficient prediction accuracy, and lack of dynamic optimization in existing technologies, significantly improving the intelligence and reliability of the water treatment process. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart of a TPT-enabled intelligent dosing treatment method based on a time-series large model according to an embodiment of the present invention.

[0015] Figure 2 This is a time-series comparison chart of COD removal rates between the present invention and traditional methods in the TPT-enabled intelligent dosing treatment method based on a time-series large model according to an embodiment of the present invention. Detailed Implementation

[0016] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0017] Figure 1 This is a flowchart of a TPT-enabled intelligent dosing method based on a large time-series model according to an embodiment of the present invention. Figure 1 As shown, the TPT-enabled intelligent dosing treatment method based on a time-series large model includes: S1: Real-time data collection of water quality indicators, flow rate, pH value, and material concentration from the wastewater treatment system is achieved through the TPT IoT system, forming a multi-source time-series dataset. At the edge, the multi-source time-series dataset undergoes pre-processing, including outlier removal and missing value completion. The processed multi-source time-series dataset is then uploaded to the cloud analysis platform via an edge-cloud collaborative transmission mechanism to ensure the real-time performance and integrity of data transmission.

[0018] This step utilizes a TPT (Total Physical Temperature) IoT sensing system deployed at the wastewater treatment site to achieve real-time monitoring of key process parameters across all dimensions. Specifically, high-precision sensor arrays are deployed at key nodes of the wastewater treatment system, such as the inlet, reaction tank, and sedimentation tank. Among them, water quality index sensors are responsible for monitoring the concentration parameters of core pollutants such as chemical oxygen demand (COD), biochemical oxygen demand (BOD), and total suspended solids (TSS). Flow sensors use electromagnetic or ultrasonic flow meters to measure the changes in water flow in each treatment unit in real time. pH sensors continuously monitor the acidity and alkalinity of the water body using the glass electrode method, and material concentration sensors monitor the actual concentration distribution of added agents such as polyaluminum chloride (PAC) and polyacrylamide (PAM).

[0019] All sensors synchronously collect data at 30-second sampling intervals according to a unified timestamp standard, forming a three-dimensional multi-source time-series dataset containing time, space, and parameter dimensions. To ensure data quality, edge computing nodes deployed at the wastewater treatment site perform real-time cleaning and preprocessing operations on the collected multi-source time-series dataset. An outlier detection algorithm based on the 3σ criterion is used to identify and remove abnormal data points caused by sensor failures, environmental interference, and other factors. At the same time, linear interpolation and moving average methods are used to intelligently fill in missing values ​​caused by data transmission interruptions or temporary sensor failures, ensuring the continuity and integrity of the data sequence.

[0020] The processed multi-source time-series datasets are efficiently uploaded through a specially designed edge-cloud collaborative transmission mechanism. This mechanism employs data compression algorithms to reduce transmission bandwidth consumption and establishes multi-path redundant transmission channels and breakpoint resume functionality. When the network connection is unstable, it can automatically switch transmission paths and resume interrupted data transmission. At the same time, a data caching mechanism is set up on the edge side to temporarily store data that was not successfully uploaded. Once the network is restored, the data is re-uploaded in batches, thereby ensuring the real-time and completeness of the data collected on-site and transmitted to the cloud analysis platform, providing high-quality basic data support for subsequent time-series predictive analysis.

[0021] S2: Input the multi-source time-series dataset into a time-series prediction model trained and optimized based on industrial scenario data, and mine the time correlation features in the multi-source time-series dataset; based on the mined time correlation features, accurately predict the water quality change trend and material change trend through the time-series prediction model, and output the optimal dosing parameter combination including dosing type parameters, dosing dosage parameters and dosing frequency parameters, and control the prediction error of the optimal dosing parameter combination within 5%.

[0022] First, the multi-source time-series dataset containing water quality index data, flow data, pH value data, and material concentration data is aligned and synchronized according to a unified timestamp. The Z-score standardization method is used to eliminate the dimensional differences between different parameters, ensuring that all types of data are processed within the same numerical range. After the data preprocessing is completed, the system divides the standardized multi-source time-series dataset into groups according to a fixed batch size, such as 64 time steps, and constructs a sliding window input sequence. Each input sample contains observations from 96 historical time points and predicted targets for 24 future time points.

[0023] During the model input phase, the time-series prediction model first converts the original numerical values ​​into high-dimensional vector representations through an embedding layer. Then, it uses a location encoding mechanism to add location information for each time step. If missing data or outliers are encountered, the model automatically calls a masking mechanism to ignore these invalid inputs. Next, a multi-head attention mechanism begins to work, identifying key temporal correlation patterns by calculating attention weight matrices between different time points and different variables. When the model processes the correlation between water quality indicators and flow data, the attention mechanism automatically focuses on the response changes of water quality indicators within 2-3 hours after the flow peak occurs. Simultaneously, the convolutional layer uses multiple one-dimensional convolutional layers of different sizes... The kernel scans the input sequence to extract local temporal pattern features such as short-term fluctuations and abrupt changes, and retains the most significant feature information through pooling operations. In the deep feature extraction process, the recurrent neural network layer selectively memorizes and forgets historical information through a gating mechanism. When a periodic change in a certain parameter is detected, the model will strengthen the memory weight of that periodic pattern. Finally, the extracted time-related features are fused and transformed through a multilayer perceptron to output a comprehensive feature vector containing trend features, periodic features, causal relationship features, and interaction features, providing a rich feature foundation for subsequent water quality change trend prediction and dosing parameter optimization.

[0024] The multi-head attention mechanism quantifies the strength of causal relationships between different variables by constructing a query-key-value triple matrix. First, the values ​​of each variable at different time points in the input multi-source time-series dataset are linearly transformed and mapped to query vector Q, key vector K, and value vector V, respectively. When the system analyzes the causal relationship between water quality indicators and flow data, flow data is used as the query variable and water quality indicators as the key variable. The original attention score is obtained by calculating the dot product of the query vector and the key vector, which reflects the strength of the impact of flow changes on water quality indicators. To identify time-lag causal relationships, the model constructs a time-shifted key matrix, matching the current query vector with key vectors from different past time points. By comparing the attention scores at different time shifts, the time delay of the strongest causal relationship is determined. For example, the flow query vector at time t and the key vector at time t-2 are compared. The highest attention score for the pH value bond vector indicates a 2-hour lag causal effect of flow rate changes on pH. In the multi-head design, each attention head is specifically responsible for capturing a particular type of causal relationship. The first head focuses on short-term direct causal relationships, the second head identifies medium-term transitive causal relationships, and the third head mines long-term cumulative causal effects. When there are complex nonlinear causal relationships between a pair of variables, different attention heads calculate the causal association strength from different perspectives and perform weighted fusion. The system normalizes the original attention scores using a softmax function to ensure that the sum of the causal association strengths among all variables is 1, thereby achieving a quantitative measurement of causal relationships. When the attention weight exceeds a preset threshold, such as 0.3, a significant causal association is considered to exist, and this weight value is used as a quantitative indicator of the causal association strength, ultimately forming a dynamically updated causal association matrix. More specifically, the causal association matrix is ​​shown in the following formula: , in, Let variable i lag variable j in time. The strength of the causal relationship below For the bulls to focus on the number of heads, Let h be the normalized weights of the h-th attention head. Let i be the query vector for the h-th head. Let variable j at time... The transpose of the h-th head key vector. The dimension of the key vector. The time decay coefficient, Let variable j at time... The h-th head value vector, For nonlinear activation parameters, For the total number of variables, To be the maximum time lag step, is the numerical stability constant.

[0025] Furthermore, in the deep feature extraction process, the recurrent neural network layer uses the gating mechanism of the LSTM (Long Short-Term Memory) network to selectively remember and forget historical information. This mechanism consists of three core components: the forget gate, the input gate, and the output gate. When new time-series data is input into the network, the forget gate first calculates the forgetting probability value through the sigmoid activation function. This value, between 0 and 1, is used to determine which historical information to discard from the cell state. For example, when water quality parameters remain stable for 48 consecutive hours, the forget gate will output a larger forgetting probability value to clear out historical data that is too old, avoiding irrelevant information from interfering with the current prediction. If a sudden change in key parameters such as flow rate or pH value is detected, the input gate will calculate the importance weight of the current input information, generate candidate update values ​​through the tanh function, and then multiply them with the output of the input gate to obtain the amount of new information to be added to the cell state. When the pH value suddenly drops from 7.2 to 6.5, the input gate will assign a higher importance weight to this change to ensure that the key information is effectively remembered.

[0026] During the information update process, the cell state combines the output of the forget gate with the cell state of the previous time step through element-wise multiplication, while adding new information filtered by the input gate, thus achieving dynamic updates of historical information. When a wastewater treatment parameter exhibits a clear diurnal cycle variation pattern, the network selectively retains data from the same time 24 hours ago as important historical reference information. The output gate is responsible for controlling which information in the cell state should be output to the hidden state. The output weight is calculated using the sigmoid function and multiplied with the tanh-activated cell state. If the current time step requires prediction of the dosage, the output gate will prioritize outputting historical treatment experience information related to the effect of the dosage, while suppressing historical data irrelevant to the current prediction task. The entire gating mechanism continuously learns and optimizes each gating parameter through the backpropagation algorithm, enabling the network to adaptively determine which historical information is most valuable for the current prediction, thereby achieving intelligent memory management and information filtering.

[0027] Furthermore, the multilayer perceptron feature fusion process employs a hierarchical transformation and progressive integration strategy. First, various temporal-related features extracted from preceding modules are standardized through the input layer to ensure that features from different sources are processed within the same numerical range. When the feature vector is input to the first hidden layer (containing 512 neurons), a linear transformation is performed on the input features using a weight matrix W1 and a bias vector b1. Then, a ReLU activation function is used to introduce non-linear characteristics, achieving initial feature fusion and abstract representation. If redundant information is detected in certain feature dimensions, the second hidden layer compresses the feature space to 256 dimensions through dimensionality reduction transformation. Simultaneously, a dropout mechanism is used to randomly zero out some neuron outputs to prevent overfitting. This layer pays particular attention to the cross-combination of different types of features, through inter-neuron... The fully connected structure learns complex nonlinear relationships between features. In the deep feature transformation stage, the third hidden layer adopts a residual connection mechanism, adding the output of the current layer to the input element-wise. When some important original feature information decays during deep propagation, the residual connection can effectively maintain the transmission of key information. At the same time, this layer stabilizes the training process and accelerates convergence through batch normalization. Finally, the output layer maps the hidden layer features into a fixed-dimensional comprehensive feature vector through linear transformation. Each dimension of this vector is processed by a sigmoid or tanh activation function to ensure that the output value is within a reasonable range. After the comprehensive feature vector is generated, its information density is significantly improved. The association patterns originally scattered in different feature types are effectively integrated to form a compact feature representation that can comprehensively represent the complex temporal dependencies of time-series data. As shown in the following equation: , in, , in, For the comprehensive feature vector, This is the first layer weight matrix. This is the weight matrix for the second layer. This is the output layer weight matrix. This is the second hidden state. For residual connectivity features, This is the first layer bias vector. This is the second layer bias vector. This is the output layer bias vector. As a time lag weight, For elements of the causal relationship matrix, For maximum time lag, The number of feature dimensions. For the input feature vector, For residual connectivity coefficients, For causal association weights, For optimal time lag, This is the element-wise multiplication operator.

[0028] Furthermore, based on the mined temporal correlation features, the time series prediction model performs accurate predictions through a sequence-to-sequence decoding mechanism. First, the comprehensive feature vector is input into the prediction decoder, which uses a multi-layer LSTM structure combined with an attention mechanism to gradually generate prediction results for future time steps through autoregression. When the model receives a comprehensive feature vector containing trend features, periodic features, causal correlation features, and interaction features, the decoder will predict the temporal evolution trajectory of key water quality indicators such as COD concentration, BOD concentration, and suspended solids concentration in the next 2-8 hours based on historical water quality change patterns and the current system state.

[0029] If the prediction indicates a risk of exceeding the standard for a certain water quality indicator, the model will simultaneously predict the material change trend, including the concentration decay rate of existing agents in the treatment tank, the duration of their effect, and the synergistic effect between different agents. By dynamically modeling the diffusion process and reaction kinetics after agent addition, the model accurately calculates the temporal and spatial distribution changes of agent concentration. In the dosing parameter optimization stage, based on the predicted water quality change trend and the target effluent standard, the model initiates a multi-objective optimization algorithm to solve for the parameters. When the system detects that the COD concentration is expected to exceed the emission standard in 4 hours, the optimization algorithm will comprehensively consider constraints such as treatment cost, agent inventory, and equipment operating status, and search for the optimal solution space using a genetic algorithm or particle swarm optimization algorithm. The system provides parameters for the types of chemicals used, such as the specific combination of polyaluminum chloride (PAC) as the main coagulant and polyacrylamide (PAM) as the coagulant aid. Dosage parameters include the precise dosage of each chemical, such as a PAC dosage of 15.3 kg per hour and a concentration ratio of 8%. It also includes parameters for the frequency of dosing, such as whether to use continuous dosing or a pulse dosing strategy every 30 minutes. The entire optimization process is iteratively improved through a reinforcement learning mechanism. When the actual treatment effect deviates from the predicted result, the model automatically adjusts the prediction algorithm parameters and the optimization objective function to ensure that the optimal combination of dosing parameters achieves a balance between minimizing costs and optimizing treatment effect while meeting water quality standards.

[0030] The multi-objective optimization algorithm employs a phased initialization and iterative solution strategy. First, the system establishes a mathematical model of the optimization problem based on current water quality predictions and target effluent standards. Water quality compliance rate, treatment cost, reagent consumption, and equipment energy consumption are the four main optimization objectives. Boundary conditions such as upper and lower limits for reagent dosage, reagent inventory constraints, and equipment processing capacity limitations are set as constraints. Once the optimization problem is constructed, the algorithm uses the NSGA-II non-dominated sorting genetic algorithm to solve it. First, an initial population of 200 individuals is randomly generated. Each individual represents a complete set of reagent dosing parameters, including reagent type selection, dosage, and dosing frequency. If the system detects that certain parameter combinations have performed well in historical operations, the algorithm adds these excellent solutions as seed individuals to the initial population to accelerate convergence. During iterative optimization, the algorithm selects parent individuals from the current population through a tournament selection mechanism, and then generates offspring individuals using simulated binary crossover and polynomial mutation operations. After the new individuals are generated, the system calls a water quality simulator to evaluate the objective function value of each solution, including predicted COD removal efficiency, total treatment cost, and reagent utilization rate.

[0031] In the core stage of multi-objective optimization, the algorithm employs a non-dominated sorting method to stratify all candidate solutions according to Pareto dominance. When two solutions are not mutually dominant, crowding distance calculation is used to maintain solution diversity, ensuring that the algorithm can find a uniformly distributed Pareto front solution set. If the Pareto front does not significantly improve in 50 consecutive iterations, the algorithm automatically adjusts the crossover and mutation probabilities to escape local optima. Simultaneously, an elitist retention strategy is introduced to ensure that the optimal solution is not lost during the evolution process. When the algorithm reaches the preset maximum number of iterations or convergence accuracy requirement, the system selects the solution that best meets the current operational needs from the Pareto optimal solution set as the final dosing parameter combination. The decision support module provides operators with multiple alternative solutions. The entire optimization process typically takes less than 30 seconds, ensuring that it meets the timeliness requirements of real-time decision-making. Specifically, the multi-objective optimization function is shown in the following equation: , in, This is the evaluation index for Pareto optimal solutions. Let be the weighting coefficient for the k-th objective, where k includes the water quality compliance rate, treatment cost, reagent consumption, and equipment energy consumption. For the k-th objective function value, Let be the expected value of the k-th objective. For non-dominated ordering decay parameters, To solve for the non-dominated ordering level of x Adjust parameters for crowding distance. For individual i, the crowding distance To constrain the penalty coefficient, The total number of decision variables, For the j-th constraint violation function, Let j be the current value of the j-th decision variable. Let j be the upper bound of the decision variable. Let be the lower bound of the j-th decision variable.

[0032] Preferably, taking the actual operation scenario of a wastewater treatment plant as an example, when the system receives multi-source time-series data at 14:30 in the afternoon showing that the influent COD concentration is 380 mg / L, ammonia nitrogen concentration is 35 mg / L, total phosphorus concentration is 4.8 mg / L, and flow rate is 950 m³ / h, the water quality prediction module, based on the causal correlation matrix... Analysis revealed a 2-hour lag correlation between COD and flow rate (correlation strength 0.73), predicting that the COD concentration would rise to 420 mg / L within the next 3 hours. If the current treatment parameters are maintained, the effluent COD will exceed the discharge standard of 35 mg / L. At this point, a multi-objective optimization algorithm automatically activates. The system establishes four objective functions when building the optimization model: minimizing the loss of water quality compliance rate. Minimize processing costs Minimize processing costs , ,in This refers to the unit price of polyaluminum chloride. This refers to the unit price of polyacrylamide. and These correspond to the respective dosages; in the constraint settings, the system limits the PAC dosage range based on the equipment's processing capacity. The range of PAM dosage is: The frequency range of addition is The algorithm operates at a rate of [number] times per hour, while considering the constraints of remaining PAC (220 kg) and PAM (38 kg) to sustain continuous operation for 6 hours. It uses NSGA-II for population initialization with 200 individuals, and performs genetic operations through tournament selection, simulated binary crossover (crossover probability 0.9), and polynomial mutation (mutation probability 0.1). The algorithm converges at generation 127, and the Pareto front solution set contains 15 non-dominated solutions. The three solutions with the highest fitness are identified by the formula [formula missing]. Calculations show that Option 1 involves a PAC dosage of 24.3 kg / h combined with a PAM dosage of 3.2 kg / h, applied every 15 minutes. =0.91, with an estimated COD removal rate of 94.2% and a total cost of 134 yuan / hour. Option 2 involves a continuous dosing mode with PAC dosage of 19.8 kg / h and PAM dosage of 2.6 kg / h. =0.89, with an estimated COD removal rate of 91.8% and a total cost of 118 yuan / hour. Option 3 involves a PAC dosage of 27.1 kg / h combined with PAM 3.8 kg / h and 1.2 kg / h of sodium hydroxide to adjust the pH value, using a continuous dosing mode. =0.87, with an estimated COD removal rate of 96.1% and a total cost of 156 yuan / hour; the system comprehensively evaluates the current operating status, which shows a clear trend of deterioration in influent water quality, strict environmental supervision requirements, and relatively sufficient reagent inventory. Ultimately, Scheme 1 is selected as the optimal combination of dosing parameters. This parameter combination can ensure that the effluent COD concentration is stably controlled below 32 mg / L, the ammonia nitrogen concentration is controlled below 4.5 mg / L, and the total phosphorus concentration is controlled below 0.4 mg / L, with an overall compliance rate of 98.5%. The optimization decision-making process takes 26.8 seconds, meeting the real-time control requirements.

[0033] S3: The optimal dosing parameter combination is converted into a dosing control command through the TPT-enabled industrial control link, and the dosing control command is sent to the automatic dosing equipment to execute the dosing operation; the treatment effect data after dosing is collected in real time, and the treatment effect data is compared and analyzed with the expected treatment target to generate parameter deviation correction values ​​for optimizing the time series prediction model.

[0034] The logic for converting the optimal dosing parameter combination into control commands employs a hierarchical mapping and protocol conversion mechanism. First, the system converts the dosing type parameters output by the optimization algorithm into specific equipment control codes via a reagent selection mapping table. When the parameter display shows polyaluminum chloride (PAC) as the selected option, the mapping logic searches the equipment configuration table to determine that PAC corresponds to dosing pump #1 and reagent storage tank #3, generating equipment start commands such as "PUMP_01_START" and valve opening commands such as "VALVE_03_OPEN," while simultaneously shutting down other equipment related to unselected reagents. If the dosing dosage parameter is 15.3 kg / h, the conversion logic calculates the corresponding pump frequency control value based on the dosing pump's characteristic curve and reagent density parameters, converts the mass flow rate to volumetric flow rate using a linear interpolation algorithm, and then maps it to... The inverter's frequency setting is, for example, 45.2Hz, and simultaneously generates a pump speed adjustment command "SET_PUMP_FREQ_45.2" which is transmitted to the inverter controller. During the conversion of the dosing frequency parameter, when the parameter is set to intermittent dosing, adding 5 minutes every 30 minutes, the timing control logic generates a periodic timed task command, including parameters such as start timestamp, duration, and repetition period, generating a timed control command similar to "TIMER_START_300s_DURATION_300s_REPEAT_1800s". If it is set to continuous dosing mode, a continuous operation command "CONTINUOUS_MODE_ENABLE" is generated. In the command encapsulation stage, the system uses the target device's communication protocol, such as Modbus, to... RTU or Profinet protocols package control parameters into standard data frames. When using the Modbus protocol, the system assembles the device address, function code, register address, and data value into a complete data packet according to the protocol specifications, and adds a CRC checksum to ensure the reliability of data transmission. If the target device supports the OPC UA protocol, the conversion logic will encapsulate the control command as a write operation to a node variable, and use variable identifiers such as "ns=2;s=PumpControl.Frequency" to precisely specify the controlled object.

[0035] First, the system precisely aligns the real-time collected treatment effect data with the expected target output by the prediction model according to the timestamp. When the COD online monitoring instrument shows an actual removal rate of 82.4% while the expected target is 88.5%, the deviation calculation module automatically calculates the relative deviation as -6.1% and judges the quality of the treatment effect based on the sign and magnitude of the deviation. If multiple water quality indicators show deviations simultaneously, the system will use a weighted comprehensive evaluation method to calculate the overall deviation index. The deviations in COD removal rate, turbidity reduction, and pH adjustment are weighted and averaged according to their impact on the final effluent quality. When the comprehensive deviation index exceeds the preset threshold of 5%, the correction algorithm will automatically start the parameter adjustment process. In the deviation cause analysis stage, the system identifies the root cause of the deviation through time-series correlation analysis and causal reasoning mechanisms. When it is found that the actual dosage of the reagent is consistent with the set value but the treatment effect is still poor, the algorithm will check the fluctuation of the influent water quality, External factors such as temperature changes and equipment operating status can influence the correction. If the influent COD concentration is detected to be 15% higher than expected, the correction logic will increase the correction magnitude of the reagent dosage accordingly. The parameter correction value generation process uses an adaptive PID control algorithm combined with a machine learning prediction mechanism. It calculates the accurate correction amount by processing the current deviation with the proportional term, eliminating the cumulative deviation with the integral term, and predicting the deviation change trend with the derivative term. When the COD removal rate deviation is -6.1%, the correction algorithm will search for successful correction examples under similar operating conditions in the historical experience database to calculate the specific correction parameters that require an increase of 2.3 kg / hour in PAC dosage and an adjustment of the dosing frequency from 30 minutes to 25 minutes. If the correction effect of a certain parameter is not good in historical data, the correction logic will activate a multi-parameter collaborative adjustment strategy to achieve a better correction effect by simultaneously adjusting the dosage and dosing time. The final output correction value can effectively improve the treatment effect deviation.

[0036] Furthermore, the real-time monitoring data is mapped one-to-one with the expected values ​​of the prediction model according to four core indicators: COD removal rate, turbidity reduction rate, pH adjustment accuracy, and reagent utilization efficiency. When the actual treatment effect differs from the expected target, the system calculates the absolute and relative deviations and marks them as "negative deviation - insufficient treatment" or "positive deviation - overtreatment" based on the direction of the deviation. Next, the system performs a severity-based judgment of the deviation. If the relative deviation is within a slight range, it is marked as a mild correction requiring only minor parameter adjustments; if it is within a moderate range, it is marked as a moderate correction requiring significant adjustments; and if it is within a severe range, it is marked as a severe correction requiring complete re-optimization. The system then triggers a correction strategy of corresponding intensity based on the deviation level. In the correction value calculation stage, the system first uses sensitivity analysis to determine the dominant influencing factors and identify which dosing parameter contributes to the current deviation. To maximize the impact, if the improvement of a certain water quality indicator is found to be mainly affected by the dosage of a specific reagent, the correction logic will prioritize adjusting that parameter, calculating the corresponding correction range through established empirical formulas and historical databases. If a single parameter correction is not expected to completely eliminate the deviation, the system will activate a multi-parameter coordinated correction mechanism. Through deviation decomposition, the overall deviation will be attributed to different causes. For example, some deviations may be attributed to insufficient dosage, requiring an increase in dosage; some deviations may be attributed to improper dosage timing, requiring an adjustment in dosage frequency; and some deviations may be attributed to reagent ratio issues, requiring optimization of concentration settings. Ultimately, a comprehensive correction parameter combination is generated, including dosage adjustment, time interval optimization, and concentration ratio correction. Each correction value is checked against constraints to ensure it does not exceed the equipment's capacity and safe operating limits, forming precise correction guidance that can be directly used for the next round of parameter optimization.

[0037] Furthermore, the collected corrected parameter combinations are added to the model update dataset as new training samples. When the corrected parameters show a systematic deviation between the actual optimal dosage and the model's prediction, the optimization mechanism activates the incremental learning module. This module adjusts the weight parameters related to dosage prediction in the neural network using the backpropagation algorithm. Specifically, the corrected true optimal parameters are used as new supervision signals to calculate the gradient of the prediction error and update the network weights. If the corrected data indicates a change in the importance of certain features, the model dynamically adjusts the attention weight distribution of the feature fusion layer. When the influence of temperature on the drug effect is found to be greater than expected, the system increases the weight coefficient of the temperature feature in the multi-head attention mechanism while correspondingly reducing the influence of other secondary features. During optimization, the system uses an experience replay mechanism to maintain an experience buffer containing historical successful correction examples. When new corrected samples are added, the optimization algorithm randomly selects historical samples from the buffer and combines them with the new samples to form a small batch of data for joint training, preventing the model from forgetting effective patterns learned in the past due to new data. If multiple consecutive corrections point to the same type of prediction deviation, the model... The optimization mechanism identifies such systematic errors and initiates structural adjustments. It enhances the model's expressive power by increasing the number of hidden layer neurons or adjusting the activation function type. When a nonlinear relationship is found to be more complex than expected, the system automatically switches from the ReLU activation function to the more complex Swish or GELU activation function. During parameter updates, the optimization algorithm employs an adaptive learning rate adjustment mechanism, dynamically adjusting the learning rate based on the confidence level of the corrected parameters and historical validation results. A larger learning rate is used to accelerate convergence when the correction effect is significant and stable, while a smaller learning rate is used to maintain model stability when the correction effect is uncertain. The entire optimization process also includes model performance validation and rollback mechanisms. The system tests the prediction accuracy of the updated model on an independent validation dataset. If the performance degradation exceeds a threshold, the optimization mechanism automatically rolls back to the previous model version and adopts a more conservative update strategy. Through this continuous optimization and self-correction mechanism, the time-series prediction model can continuously learn from actual operating experience, gradually improving its adaptability to complex operating conditions and prediction accuracy, ultimately achieving deep integration and collaborative optimization between the prediction model and the actual control effect.

[0038] S4: Feed the processing effect data and parameter deviation correction values ​​back to the time series prediction model for automatic iterative optimization, update the time-related feature weights and prediction algorithm parameters in the model; when complex operating condition fluctuations are detected, manual intervention is supported to adjust the optimal dosing parameter combination, forming a dynamic closed-loop control mechanism of "TPT data acquisition - time series model prediction - control command issuance - effect data feedback".

[0039] The feedback optimization mechanism for treatment effect data and parameter deviation correction values ​​adopts a multi-level adaptive learning architecture. First, the system establishes a dedicated feedback data processing module to preprocess and quality-verify the treatment effect data from on-site monitoring equipment. Outliers and noise data are removed using data cleaning algorithms to ensure the reliability of the feedback information. When the treatment effect data shows that the effluent COD concentration is consistently lower than the expected target, the feedback processing module automatically marks this type of positive deviation data and extracts its corresponding process conditions and dosing parameters as excellent samples. Simultaneously, the parameter deviation correction values ​​are categorized and organized according to dimensions such as correction type, correction magnitude, and correction effect, establishing a structured feedback database. In the automatic iterative optimization phase, after receiving feedback data, the time-series prediction model starts an online learning algorithm, integrating new feedback samples into the existing training dataset through incremental learning. When the system detects a continuous decline in prediction accuracy within a certain time period, the optimization algorithm focuses on analyzing the feedback data for that period to identify the key factors causing prediction deviations. If it is found that the inaccurate prediction is due to sudden changes in influent water quality, the model automatically strengthens the learning weights for influent water quality fluctuation characteristics by adjusting... The attention mechanism uses weight distribution to enhance the influence of relevant features. The update process of time-related feature weights adopts a strategy combining gradient backpropagation and feature importance assessment. The system dynamically adjusts the weight coefficients of each time-related feature by calculating its contribution to the final prediction error. When the influence of the historical 24-hour water quality change trend on the current prediction is found to be greater than the originally set weight, the optimization algorithm will correspondingly increase the weight of long-term trend features and decrease the weight of short-term fluctuation features. The entire weight adjustment process uses the exponential moving average method to smooth weight changes and avoid drastic weight fluctuations caused by a single abnormal data. The update of prediction algorithm parameters involves multiple aspects such as the weight matrix, bias vector, and activation function parameters of each layer of the neural network. The system uses an adaptive optimizer such as AdamW to automatically adjust the learning rate and parameter update magnitude. When feedback data indicates that the prediction system is systematically low under certain working conditions, the optimization algorithm will focus on adjusting the weights and bias parameters of the output layer to correct this systematic error. If it is found that the model's ability to fit certain nonlinear relationships is insufficient, the system will dynamically adjust the structural parameters or activation function type of the hidden layer to enhance the model's expressive power.

[0040] The system establishes a multi-level early warning and manual intervention mechanism for detecting fluctuations in complex operating conditions. It identifies complex operating conditions by real-time monitoring statistical indicators such as the coefficient of variation, trend change rate, and frequency of outliers for key process parameters. When it detects that the influent COD concentration fluctuates beyond the normal range within a short period, the pH value changes drastically, or multiple water quality indicators deviate from the normal range simultaneously, the system automatically identifies it as a complex operating condition and activates the early warning mechanism, sending alarm information to the operator. During the manual intervention and adjustment phase, the system provides operators with an intuitive human-machine interface, displaying the current optimal dosing parameter combination, prediction confidence level, historical treatment effect trends, and other relevant information. The system allows operators to manually adjust parameters such as the type, dosage, and frequency of chemicals based on actual site conditions and professional experience. When operators discover that a special pollutant requires unconventional treatment agents, they can select a backup agent and set the corresponding dosage parameters through the interface. The system will automatically record these manual adjustments and store them in the knowledge base as experience for handling special conditions. The manual intervention process also includes parameter rationality checks and safety verification functions. If the parameters entered by the operator may cause equipment overload or violate safety procedures, the system will automatically pop up a warning message and suggest a reasonable parameter range to ensure the safety and effectiveness of manual adjustments.

[0041] The construction of the dynamic closed-loop control mechanism realizes full-process automation and intelligent management from data acquisition to control execution. The entire closed-loop system starts with TPT data acquisition and collects multi-dimensional data in real time, including water quality parameters, equipment operating status, and environmental conditions, through sensors and online monitoring equipment deployed at various key nodes of the treatment process. The data acquisition frequency is set differently according to the parameter type and importance. Key water quality indicators such as COD and BOD are collected at high frequencies, while general operating parameters are collected at medium and low frequencies. All collected data is transmitted to the central data processing platform via industrial Ethernet or wireless communication. The time series model prediction stage receives the pre-processed real-time data and combines it with historical data and the current system status to predict future water quality change trends and calculate optimal dosing parameters. The prediction process considers various uncertainties and constraints. The system outputs prediction results with confidence intervals and corresponding dosing control strategies. During the control command issuance phase, the abstract parameters output by the prediction model are converted into specific equipment control commands, which are then sent to the automated dosing equipment in the field via standardized industrial communication protocols, achieving precise dosing control. In the effect data feedback phase, the system continuously monitors the treatment effect after dosing, compares and analyzes the actual effect data with the expected target, generates performance evaluation reports and parameter optimization suggestions, forming a complete data closed loop. This closed-loop control mechanism possesses self-learning, self-adaptive, and self-optimizing characteristics, enabling continuous improvement of control strategies based on system operating experience, adapting to different operating conditions and treatment needs. Through continuous feedback optimization, it achieves gradual improvement in control accuracy and treatment effect, ultimately constructing a highly intelligent and automated water treatment dosing control system.

[0042] In summary, the TPT-enabled intelligent dosing method based on a time-series large model, as described in this invention, is elucidated. It collects multi-dimensional operating parameters in real time through a TPT IoT system, forming a multi-source time-series dataset. After preprocessing at the edge, the dataset is transmitted to a cloud analysis platform. A time-series prediction model is used to mine temporal correlation features, predict water quality change trends, and output the optimal combination of dosing parameters. This enables accurate prediction and efficient control of complex water quality fluctuations, improving the precision and efficiency of dosing treatment. Through a closed-loop feedback optimization mechanism, dosing parameters can be dynamically corrected, further enhancing the system's stability and responsiveness. This effectively solves the problems of low data processing efficiency, insufficient prediction accuracy, and lack of dynamic optimization in existing technologies, significantly improving the intelligence and reliability of the water treatment process.

[0043] Here, those skilled in the art will understand that the specific operations of each step in the TPT-enabled intelligent dosing system based on the time-series large model have been referenced above. Figure 1 and Figure 2 The TPT-enabled intelligent dosing treatment method based on a time-series large model has been described in detail, and therefore, its repeated description will be omitted.

[0044] In summary, the TPT-enabled intelligent dosing system based on a time-series large model, as described in this invention, is explained. It collects multi-dimensional operating parameters in real time through a TPT IoT system, forming a multi-source time-series dataset. After preprocessing at the edge, the dataset is transmitted to a cloud analysis platform. The system utilizes a time-series prediction model to mine temporal correlation features, predict water quality change trends, and output the optimal combination of dosing parameters. This enables accurate prediction and efficient control of complex water quality fluctuations, improving the precision and efficiency of dosing treatment. Through a closed-loop feedback optimization mechanism, dosing parameters can be dynamically corrected, further enhancing the system's stability and responsiveness. This effectively solves the problems of low data processing efficiency, insufficient prediction accuracy, and lack of dynamic optimization in existing technologies, significantly improving the intelligence and reliability of the water treatment process.

Claims

1. A time-series large-scale model-based TPT-enabled intelligent dosing treatment method, characterized in that, include: The TPT IoT system collects multi-dimensional operating parameters in real time to form a multi-source time-series dataset, which is then preprocessed at the edge and transmitted to the cloud analysis platform through edge-cloud collaboration. The multi-source time-series dataset is input into the time-series prediction model to mine time-related features, predict water quality change trends based on the mined features, and output the optimal combination of dosing parameters. The optimal combination of dosing parameters is converted into control commands and sent to the automatic dosing equipment. Real-time data on the treatment effect after dosing is collected and compared with the expected target to generate parameter deviation correction values. The processing effect data and the deviation correction value are fed back to the time series prediction model for iterative optimization, updating the time-related feature weights and prediction parameters.

2. The TPT-enabled intelligent dosing method based on a large time-series model according to claim 1, characterized in that, Mining the time-related features includes the following steps: Multi-source time-series data are processed in layers by setting sliding windows with different time granularities; By combining multi-head attention mechanism to calculate dynamic correlation weights and time delay features between variables, and using convolutional neural network to extract local time patterns and identify periodic patterns through frequency domain analysis; Ultimately, causal relationship analysis and mutual information calculation are integrated to comprehensively uncover the complex temporal correlation features in time series data.

3. The TPT-enabled intelligent dosing method based on a large time-series model according to claim 2, characterized in that, The multi-head attention mechanism identifies causal relationships and time delays between variables by mapping time-series data into query-key-value triples and calculating attention scores at different time offsets. Each attention head captures short-term direct, medium-term transmission, and long-term cumulative causal effects, and finally generates a quantified causal relationship matrix through softmax normalization.

4. The TPT-enabled intelligent dosing method based on a large time-series model according to claim 3, characterized in that, The optimal dosing parameter combination is output by solving the Pareto optimal solution set through a multi-objective optimization algorithm. Combined with the current operating conditions and constraints, the analytic hierarchy process is used to comprehensively evaluate and output a structured optimal dosing parameter combination that includes specific drug types, precise dosage, and detailed dosing frequency.

5. The TPT-enabled intelligent dosing method based on a large time-series model according to claim 4, characterized in that, The multi-objective optimization algorithm includes phased initialization and iterative solution; The phased initialization includes taking water quality compliance rate, treatment cost, chemical consumption and equipment energy consumption as the main optimization objectives, while setting upper and lower limits for chemical dosage, chemical inventory constraints and equipment treatment capacity limits as constraints. The iterative solution employs a non-dominated sorting method to stratify all candidate solutions according to Pareto dominance.

6. The TPT-enabled intelligent dosing method based on a large time-series model according to claim 5, characterized in that, The multi-objective optimization algorithm is shown in the following equation: , in, This is the evaluation index for Pareto optimal solutions. Let be the weighting coefficient for the k-th objective, where k includes the water quality compliance rate, treatment cost, reagent consumption, and equipment energy consumption. For the k-th objective function value, Let be the expected value of the k-th objective. For non-dominated ordering decay parameters, To solve for the non-dominated ordering level of x Adjust parameters for crowding distance. For individual i, the crowding distance To constrain the penalty coefficient, The total number of decision variables, For the j-th constraint violation function, Let j be the current value of the j-th decision variable. Let j be the upper bound of the decision variable. Let be the lower bound of the j-th decision variable.

7. The TPT-enabled intelligent dosing method based on a large time-series model according to claim 6, characterized in that, The optimal combination of dosing parameters is converted into control commands. The dosing type parameters are converted into equipment start-stop commands through the dosing selection mapping table. The dosing dosage parameters are converted into inverter frequency setpoints based on the pump characteristic curve and the dosing density. The dosing frequency parameters are converted into timing control commands. Finally, all control parameters are encapsulated into standard data frames and checksums are added according to the communication protocol of the target equipment.

8. A TPT-enabled intelligent dosing system based on a time-series large model, characterized in that, include: The data acquisition and transmission module is used to collect multi-dimensional operating parameters in real time through the TPT IoT system, form a multi-source time-series dataset, perform preprocessing on the edge side, and transmit it to the cloud analysis platform through edge-cloud collaboration. The intelligent prediction and optimization module is used to input multi-source time-series datasets into the time-series prediction model, mine time-related features, predict water quality change trends based on the mined features, and output the optimal combination of dosing parameters. The control execution monitoring module is used to convert the optimal combination of dosing parameters into control commands and send them to the automatic dosing equipment. It also collects the treatment effect data after dosing in real time and compares it with the expected target to generate parameter deviation correction values. The feedback learning optimization module is used to feed back the processing effect data and bias correction values ​​to the time series prediction model for iterative optimization, and update the weights of time-related features and prediction parameters.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the TPT-enabled intelligent dosing method based on a time-series large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the TPT-enabled intelligent dosing treatment method based on the time-series large model as described in any one of claims 1 to 7.

Images