Drinking water automation equipment production scheduling optimization control method and system

Abstract

The present invention provides a production scheduling optimization control method and system for drinking water automation equipment, which relates to the field of control technology, including obtaining real-time production data and historical production data of drinking water automation equipment, and using a machine learning algorithm to establish a drinking water production prediction model based on the acquired real-time production data and historical production data; formulating a preliminary production plan according to the output result of the drinking water production prediction model, combined with current production orders and inventory conditions, including the start-up time, production batches and product types of each production line; then using a multi-objective optimization algorithm to optimize the preliminary production plan with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs, to obtain an optimized production scheduling plan; and issuing the optimized production scheduling plan to the control system of each drinking water automation equipment to achieve automated control of the production process.

Description

Technical Field

[0001] The invention relates to control technology, and in particular to a production scheduling optimization control method and system for drinking water automation equipment. Background Art

[0002] As a key link in ensuring the production and supply of drinking water, the production scheduling and optimization control of drinking water automation equipment directly affect the quality of drinking water, production efficiency and the economic benefits of the enterprise. However, the production scheduling and control of drinking water automation equipment still face some challenges:

[0003] Complexity of the production process: Drinking water production involves multiple links such as water collection, purification, disinfection, and storage, each of which contains multiple process parameters and control variables. Traditional manual scheduling methods are difficult to fully consider the mutual influence between each link, resulting in low overall production efficiency.

[0004] Demand volatility: There are obvious seasonal and time-based fluctuations in the demand for drinking water, which places high demands on the flexibility of production scheduling. Fixed production plans are difficult to adapt to rapid changes in demand, and can easily lead to waste of production capacity or insufficient supply. Energy consumption issues: The energy consumption in the production process of drinking water is relatively large, especially in the operation of water pumps and water treatment equipment. How to minimize energy consumption while ensuring water quality is an important challenge currently faced. Water quality control accuracy: The quality of drinking water is directly related to public health. Traditional control methods are difficult to maintain stable effluent water quality when faced with problems such as fluctuations in water quality and accuracy of chemical dosing. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for optimizing the production scheduling of drinking water automation equipment, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides drinking water automation equipment production scheduling optimization control method, including:

[0008] Obtain real-time production data and historical production data of drinking water automation equipment, where real-time production data includes raw water quality, equipment operating status and production energy consumption, and historical production data includes past production orders, equipment maintenance records and seasonal demand fluctuations; based on the acquired real-time production data and historical production data, use machine learning algorithms to establish a drinking water production prediction model, which is used to predict production demand and equipment performance within a certain period of time in the future;

[0009] According to the output results of the drinking water production prediction model, combined with the current production orders and inventory conditions, a preliminary production plan is formulated, including the start-up time of each production line, production batches and product types; then, a multi-objective optimization algorithm is used to optimize the preliminary production plan with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs, to obtain an optimized production scheduling plan; the optimized production scheduling plan is sent to the control system of each drinking water automation equipment to realize automated control of the production process;

[0010] During the production process, the operating status and production indicators of each equipment are monitored in real time. When an abnormal situation is detected, the preset emergency response process is initiated, and the production scheduling plan is dynamically adjusted according to the actual situation. At the same time, data from the entire production process is collected, including raw material usage, production efficiency, product quality and energy consumption indicators. This data is fed back into the drinking water production prediction model, and the parameters of the drinking water production prediction model are continuously optimized and updated to improve the prediction accuracy.

[0011] In an optional embodiment,

[0012] Based on the real-time production data and historical production data obtained, a machine learning algorithm is used to establish a drinking water production prediction model. The model is used to predict production demand and equipment performance in a certain period of time in the future, including:

[0013] Preprocessing the real-time production data and the historical production data includes decomposing the original time series data into trend terms, seasonal terms and residual terms using an improved seasonality and trend decomposition algorithm, wherein the improved seasonality and trend decomposition algorithm uses an adaptive window size, and the adaptive window size is determined according to the seasonal cycle and the sequence length;

[0014] Performing feature engineering on the preprocessed data, including extracting time features, generating dynamic lag features, constructing cross features, and extracting Fourier features, wherein the dynamic lag features automatically select the most relevant lag period through autocorrelation analysis;

[0015] Construct a multi-model integration framework for a drinking water production prediction model, including a base learner layer, an intermediate integration layer, and a final integration layer, wherein the base learner layer includes a long short-term memory network, a gated recurrent unit, a prophet model, and an XGBoost algorithm, the intermediate integration layer uses a dynamic weight allocation mechanism to dynamically adjust the weights according to the performance of each model on a validation set, and the final integration layer uses a stacked generalization method to use an elastic network as a meta-learner; the constructed multi-model integration framework is used to predict production demand and equipment performance within a certain period of time in the future;

[0016] An incremental learning strategy is adopted for online learning and model updating, including the use of a concept drift detection algorithm to detect changes in data distribution and triggering local model updates when significant changes are detected. The update process uses progressive fine-tuning techniques.

[0017] In an optional embodiment,

[0018] The preliminary production plan is optimized by using a multi-objective optimization algorithm with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs. The optimized production scheduling scheme includes:

[0019] Construct a multi-objective optimization model, taking minimization of energy consumption, maximization of capacity utilization and minimization of production costs as optimization objectives, while considering total demand satisfaction and capacity limitations as constraints;

[0020] Initialize the particle swarm, where each particle represents a production scheduling plan;

[0021] An adaptive inertia weight strategy is used to update particle velocity, where the inertia weight changes nonlinearly with the number of iterations to balance global exploration and local development capabilities;

[0022] Using a crowding-based leader selection strategy, the crowding of non-dominated solutions is calculated, and the leader is selected through a tournament selection method to maintain the diversity of solutions;

[0023] The ε-constraint processing mechanism is introduced to compare and update the particles according to the degree of constraint violation, where the ε value gradually decreases during the iteration process.

[0024] An adaptive mutation operator based on Gaussian distribution is applied to enhance the local search capability of the algorithm, and the mutation intensity is adaptively adjusted with the number of iterations;

[0025] Iteratively update the particle position and velocity until the termination condition is reached;

[0026] Generate a Pareto optimal solution set, use parallel coordinates and radar charts to visualize the Pareto frontier, and obtain the optimized production scheduling plan.

[0027] In an optional embodiment,

[0028] The method further comprises:

[0029] The fuzzy analytic hierarchy process is used to construct a decision support system and calculate the weights of different objectives;

[0030] Use the weighted Chebyshev method to select the final production scheduling solution from the Pareto optimal solution set;

[0031] Output an optimized production scheduling plan that strikes a balance between minimizing energy consumption, maximizing capacity utilization, and minimizing production costs;

[0032] Among them, a fuzzy judgment matrix reflecting the relative importance of multiple optimization objectives is established;

[0033] Calculating a fuzzy comprehensive evaluation value based on the fuzzy judgment matrix; calculating the possibility of a fuzzy number according to the fuzzy comprehensive evaluation value; calculating a weight vector using the possibility of the fuzzy number, wherein each component of the weight vector corresponds to a weight of an optimization target;

[0034] Obtain a Pareto optimal solution set, wherein each solution in the Pareto optimal solution set corresponds to a set of objective function values; for each solution in the Pareto optimal solution set, calculate the difference between each objective function value and its ideal value; divide the difference by the difference between the ideal value and the maximum value of the corresponding objective function to obtain a normalized difference; multiply the normalized difference by the weight of the corresponding optimization target to obtain a weighted normalized difference;

[0035] For each Pareto optimal solution, the maximum value of the weighted normalized difference of all its optimization objectives is taken as the evaluation index of the solution; the Pareto optimal solution with the minimum evaluation index is selected as the final decision result.

[0036] In an optional embodiment,

[0037] The optimized production scheduling plan is sent to the control system of each drinking water automation equipment to realize the automated control of the production process, including:

[0038] Obtain optimized production scheduling plan;

[0039] Convert the production scheduling plan into standardized digital instructions, securely encrypt the converted digital instructions, and use X.509 certificate authentication and AES-256 encryption algorithm to ensure data transmission security;

[0040] Transmit the encrypted digital instructions to the distributed control system via the OPC UA protocol;

[0041] Parse the received digital instructions in the distributed control system and dynamically adjust the control strategy, including updating PID control parameters and fuzzy control rules;

[0042] Send the parsed control instructions to the control systems of each drinking water automation equipment;

[0043] In the device control system, intelligent actuators supporting the HART protocol are used to execute control instructions to achieve precise device-level control. Edge computing nodes are deployed, and control algorithms are encapsulated using Docker container technology to achieve real-time control and flexible algorithm upgrades at the device layer.

[0044] In an optional embodiment,

[0045] Parse the received digital instructions in the distributed control system and dynamically adjust the control strategy, including updating PID control parameters and fuzzy control rules:

[0046] Receiving a digital instruction, wherein the digital instruction is semantically encapsulated in JSON-LD format;

[0047] Parsing the digital instruction using a parser based on a finite state machine, wherein the finite state machine comprises five states: initialization, parsing header, parsing body, verification and execution;

[0048] During the parsing process, the try-catch mechanism is used to capture and classify parsing exceptions;

[0049] According to the analysis results, determine the control strategy that needs to be updated, including PID control parameters and fuzzy control rules;

[0050] For the update of PID control parameters, an adaptive adjustment method based on reinforcement learning is adopted, including:

[0051] Define the state space and take the system error, error change rate and integrated error as state variables;

[0052] Design a reward function that takes into account steady-state error, dynamic response, and control energy consumption;

[0053] Use the Q-learning algorithm to update the Q value and adopt the ε-greedy strategy for action selection;

[0054] Introduce a priority experience replay mechanism to perform importance sampling on samples according to the error size;

[0055] For the update of fuzzy control rules, an adaptive adjustment method based on a combination of genetic algorithm and fuzzy neural network is adopted, including:

[0056] Encode fuzzy rules into chromosomes, including membership function parameters and rule weights;

[0057] Design the fitness function, taking into account the mean square error and the complexity of the rule base;

[0058] Perform genetic operations including roulette wheel and elite-preserving selection, arithmetic crossover, and Gaussian mutation;

[0059] The fuzzy rules optimized by genetic algorithm are used as the initial structure of fuzzy neural network;

[0060] The optimized control parameters and rules are applied to the distributed control system to achieve dynamic adjustment of the control strategy.

[0061] According to a second aspect of the embodiments of the present invention,

[0062] Provide drinking water automation equipment production scheduling optimization control system, including:

[0063] The first unit is used to obtain real-time production data and historical production data of drinking water automation equipment, where the real-time production data includes raw water quality, equipment operating status and production energy consumption, and the historical production data includes past production orders, equipment maintenance records and seasonal demand fluctuations; based on the acquired real-time production data and historical production data, a machine learning algorithm is used to establish a drinking water production prediction model, which is used to predict production demand and equipment performance within a certain period of time in the future;

[0064] The second unit is used to formulate a preliminary production plan according to the output result of the drinking water production prediction model, combined with the current production orders and inventory conditions, including the start-up time of each production line, production batches and product types; then use a multi-objective optimization algorithm to optimize the preliminary production plan with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs, and obtain an optimized production scheduling plan; the optimized production scheduling plan is sent to the control system of each drinking water automation equipment to realize automated control of the production process;

[0065] The third unit is used to monitor the operating status and production indicators of each equipment in real time during the production process. When an abnormal situation is detected, the preset emergency response process is initiated, and the production scheduling plan is dynamically adjusted according to the actual situation. At the same time, data from the entire production process is collected, including raw material usage, production efficiency, product quality and energy consumption indicators, and these data are fed back to the drinking water production prediction model to continuously optimize and update the parameters of the drinking water production prediction model and improve the prediction accuracy.

[0066] According to a third aspect of the embodiments of the present invention,

[0067] An electronic device is provided, comprising:

[0068] processor;

[0069] a memory for storing processor-executable instructions;

[0070] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0071] According to a fourth aspect of the embodiments of the present invention,

[0072] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0073] Through the incremental learning strategy, the multi-model integration framework can timely adapt to changes in data distribution and maintain the stability and accuracy of prediction performance. In practical applications, this method significantly improves the model's response speed to emergencies (such as equipment failure, water quality abnormalities, etc.), providing strong support for the intelligent management of drinking water production processes.

[0074] This solution realizes intelligent control of distributed control systems through the analysis of digital instructions and the adjustment of dynamic control strategies. The solution uses reinforcement learning to adaptively adjust PID parameters, genetic algorithms and fuzzy neural networks to optimize fuzzy control rules, thereby improving the control accuracy and response speed of the system. Through continuous learning and adjustment, the system can maintain efficient and stable operation in complex industrial environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A schematic diagram of a flow chart of a method for optimizing production scheduling of drinking water automation equipment according to an embodiment of the present invention;

[0076] Figure 2 It is a structural schematic diagram of a production scheduling optimization control system for drinking water automation equipment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0079] Figure 1 FIG. 1 is a flow chart of a method for optimizing production scheduling of drinking water automation equipment according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0080] S101. Obtaining real-time production data and historical production data of drinking water automation equipment, wherein the real-time production data includes raw water quality, equipment operating status and production energy consumption, and the historical production data includes past production orders, equipment maintenance records and seasonal demand fluctuations; based on the acquired real-time production data and historical production data, using a machine learning algorithm to establish a drinking water production prediction model, which is used to predict production demand and equipment performance within a certain period of time in the future;

[0081] S102. Formulate a preliminary production plan based on the output of the drinking water production prediction model, combined with current production orders and inventory conditions, including the start-up time of each production line, production batches and product types; then optimize the preliminary production plan using a multi-objective optimization algorithm with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs, and obtain an optimized production scheduling plan; send the optimized production scheduling plan to the control system of each drinking water automation equipment to achieve automated control of the production process;

[0082] S103. During the production process, the operating status and production indicators of each device are monitored in real time. When an abnormal situation is detected, the preset emergency response process is initiated, and the production scheduling plan is dynamically adjusted according to the actual situation. At the same time, data from the entire production process is collected, including raw material usage, production efficiency, product quality and energy consumption indicators, and these data are fed back to the drinking water production prediction model, and the parameters of the drinking water production prediction model are continuously optimized and updated to improve the prediction accuracy.

[0083] In an optional embodiment,

[0084] Based on the real-time production data and historical production data obtained, a machine learning algorithm is used to establish a drinking water production prediction model. The model is used to predict production demand and equipment performance in a certain period of time in the future, including:

[0085] Preprocessing the real-time production data and the historical production data includes decomposing the original time series data into trend terms, seasonal terms and residual terms using an improved seasonality and trend decomposition algorithm, wherein the improved seasonality and trend decomposition algorithm uses an adaptive window size, and the adaptive window size is determined according to the seasonal cycle and the sequence length;

[0086] Performing feature engineering on the preprocessed data, including extracting time features, generating dynamic lag features, constructing cross features, and extracting Fourier features, wherein the dynamic lag features automatically select the most relevant lag period through autocorrelation analysis;

[0087] Construct a multi-model integration framework for a drinking water production prediction model, including a base learner layer, an intermediate integration layer, and a final integration layer, wherein the base learner layer includes a long short-term memory network, a gated recurrent unit, a prophet model, and an XGBoost algorithm, the intermediate integration layer uses a dynamic weight allocation mechanism to dynamically adjust the weights according to the performance of each model on a validation set, and the final integration layer uses a stacked generalization method to use an elastic network as a meta-learner; the constructed multi-model integration framework is used to predict production demand and equipment performance within a certain period of time in the future;

[0088] An incremental learning strategy is adopted for online learning and model updating, including the use of a concept drift detection algorithm to detect changes in data distribution and triggering local model updates when significant changes are detected. The update process uses progressive fine-tuning techniques.

[0089] Exemplarily, this application proposes an improved seasonality and trend decomposition algorithm, the core innovation of which is the use of adaptive window size. The key to this algorithm is to accurately capture multi-scale periodic patterns in time series data while adapting to data sequences of different lengths.

[0090] The adaptive window size is determined as follows: First, the autocorrelation function (ACF) of the time series is calculated, and the maximum lag period is set to 1 / 4 of the original series length. This setting ensures a sufficiently long analysis period while avoiding the problem of excessive computational complexity. Next, we analyze the periodic peaks of the ACF and identify significant periodic patterns. In practice, it is found that drinking water production data usually exhibits multiple periodicities, such as 24 hours, 7 days, and 30 days.

[0091] In order to accurately identify these cycles, a peak detection-based approach was used. Specifically, the ACF curve was first smoothed using a Savitzky-Golay filter with a window size set to 1% of the ACF length and a polynomial order of 3. This step effectively removed the noise in the ACF, making the true periodic peaks more obvious. Then, the find_peaks function in the scipy library was used to detect local maxima, setting the minimum peak height to 0.1 and the minimum peak distance to 24 (corresponding to the daily cycle).

[0092] After identifying multiple significant cycles, the largest cycle is selected as the window size. This approach ensures that the decomposition algorithm can capture the longest periodic pattern while also including shorter cycles. For example, in the experiment, for the water production data of a water plant, three significant cycles of 24 hours, 7 days, and 30 days were found, and 30 days (8640 data points) were finally selected as the window size.

[0093] The original time series is decomposed using a determined adaptive window size. The trend term is extracted using the centralized moving average method, and the moving window size is the adaptive window size. This method can effectively capture long-term trends while retaining sufficient flexibility to adapt to different data characteristics.

[0094] The extraction of seasonal terms is more complicated. First, the trend term is subtracted from the original sequence, and then the average pattern of each seasonal cycle is calculated. The innovation here is that not only the major cycles (such as 30 days) but also the minor cycles (such as 24 hours and 7 days) are considered. Specifically, the 30-day periodic pattern is first extracted, then the 7-day periodic pattern is extracted from the remaining sequence, and finally the 24-hour periodic pattern is extracted. This multi-scale seasonal extraction method can more comprehensively capture the periodic changes in the data.

[0095] The residual term is the remainder of the original series after subtracting the trend term and seasonal term. It is worth noting that further analysis of the residual term, including white noise test and autocorrelation analysis, was performed to ensure that no important patterns were missed.

[0096] This improved seasonality and trend decomposition algorithm performs well in practical applications. Taking the daily water production data of a water plant as an example, using the traditional STL decomposition method with a fixed window size (30 days), there is still an obvious periodic pattern in the residual term, and the mean absolute error (MAE) is 245 cubic meters / day. However, using our improved algorithm, the residual term is closer to white noise, and the MAE is reduced to 187 cubic meters / day, an improvement of 23.7%. This result fully demonstrates the advantage of the adaptive window method in capturing complex time series patterns.

[0097] In the feature engineering process, the generation of dynamic lagged features is a key link. Traditional methods usually use fixed lag periods, such as the first 1 hour, the first 24 hours, etc. Although this method is simple, it often cannot fully capture the long-term dependencies in the data. This study proposes a dynamic lagged feature generation method based on autocorrelation analysis, which can automatically select the most relevant lag period, thereby improving the performance of the prediction model.

[0098] First, the autocorrelation function (ACF) of the target variable (such as water production) is calculated. Considering the characteristics of drinking water production, the maximum lag period is set to 7 days (2016 data points, because the data collection frequency is 5 minutes). This setting is sufficient to capture the intra-day, intra-week and inter-week dependencies. The ACF is calculated using the correlate function of the numpy library, which is implemented by the following formula:

[0099] ACF(k) = [(X_t - μ)(X_{t+k} - μ)] / σ^2;

[0100] Among them, k is the lag period, X_t is the time series, μ is the sequence mean, σ^2 is the sequence variance, and t represents time.

[0101] The lags with an absolute value of correlation coefficient greater than 0.3 are selected as candidates. This threshold is determined after many experiments and can strike a good balance between retaining important information and controlling the number of features. In practice, it is found that this threshold usually selects 10 to 20 lags, which can provide enough information without causing the feature dimension to be too high.

[0102] For the selected lag period, the present application generates the corresponding lag features. The innovation here is that we not only generate a single lag value, but also consider the statistical characteristics within the lag period. Specifically, for each selected lag period k, the following features are generated:

[0103] - The value after k periods: X_{tk};

[0104] - The average value within the k-period lag: mean(X_{tk}, ..., X_t);

[0105] - Standard deviation within lag k: std(X_{tk}, ..., X_t);

[0106] - Maximum value within lag k: max(X_{tk}, ..., X_t);

[0107] - Minimum value within lag k: min(X_{tk}, ..., X_t);

[0108] This multi-dimensional lag feature can capture the information of historical data more comprehensively.

[0109] In order to evaluate the effect of dynamic lag features, they were compared with fixed lag features (1 hour, 24 hours, and 7 days). In an XGBoost-based forecasting model, the use of dynamic lag features reduced the root mean square error (RMSE) from the original 235 cubic meters / hour to 198 cubic meters / hour, an improvement of 15.7%. This result fully demonstrates the advantage of dynamic lag features in capturing complex time dependencies.

[0110] This application constructs a three-layer multi-model integration framework, including a base learner layer, an intermediate integration layer, and a final integration layer. This hierarchical structure can fully utilize the advantages of different models to improve the accuracy and robustness of predictions.

[0111] At the base learner layer, four different types of models were selected: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Prophet, and XGBoost. These models have their own characteristics and can capture different features of time series data.

[0112] The specific implementation of the LSTM network adopts a bidirectional LSTM structure to make full use of the contextual information of the sequence. The network structure includes an embedding layer, two bidirectional LSTM layers, and a fully connected output layer. The dimension of the embedding layer is set to 64, and each LSTM layer contains 128 hidden units. In order to prevent overfitting, a dropout layer is added between the LSTM layers, and the dropout rate is set to 0.3. The model is trained using the Adam optimizer, the initial value of the learning rate is set to 0.001, and the learning rate decay strategy is adopted, decaying by 10% every 50 epochs.

[0113] The structure of the GRU network is similar to that of the LSTM, but GRU units are used instead of LSTM units. The number of GRU layers and hidden units are kept the same as those of the LSTM for fair comparison. The other hyperparameter settings of the GRU network are the same as those of the LSTM.

[0114] The Prophet model is a model suitable for multi-seasonal time series forecasting. In this application, the Prophet model is customized to adapt to the characteristics of drinking water production. Specifically, three seasonal components are set: daily cycle, weekly cycle and monthly cycle. At the same time, additional regression variables such as weather factors and holiday effects are introduced to improve the predictive ability of the model.

[0115] As an efficient ensemble learning method, the XGBoost algorithm also played an important role in this study. The hyperparameters of XGBoost were carefully tuned, and the main parameters finally determined are as follows: the maximum tree depth is 8, the learning rate is 0.05, the subsampling rate is 0.8, and the feature sampling rate is 0.7. In addition, we also adopted an early stopping strategy and set the patience to 50 to prevent overfitting.

[0116] The intermediate integration layer uses a dynamic weight allocation mechanism to dynamically adjust the weights according to the performance of each model on the validation set. Specifically, a weight update algorithm based on the exponential moving average (EMA) is designed. For each prediction cycle (for example, every 24 hours), we calculate the root mean square error (RMSE) of each base learner in the recent period. Then, we take the inverse of the RMSE and normalize it to get the initial weight.

[0117] The final integration layer uses the stacking method and uses the elastic network as a meta-learner. The elastic network combines the advantages of L1 and L2 regularization and can achieve a good balance between feature selection and model stability. A grid search was performed on the hyperparameters of the elastic network, and α (L1 scale) was finally determined to be 0.5 and λ (regularization strength) to be 0.001.

[0118] In order to adapt to the possible changes in data distribution in the drinking water production process, an incremental learning strategy is adopted for online learning and model updating, which includes two key steps: concept drift detection and local model update.

[0119] Concept drift detection uses the improved ADWIN (Adaptive Windowing) algorithm. Compared with the original ADWIN algorithm, the improved version introduces multiple detection windows to adapt to concept drifts of different time scales. Specifically, three sliding windows of different sizes are maintained: short-term window (24 hours), medium-term window (7 days) and long-term window (30 days). Each window performs mean change detection independently.

[0120] In order to improve the reliability of detection, a confidence threshold is also introduced. The concept drift alarm is triggered only when the detected mean difference exceeds the preset threshold and lasts for a certain period of time (for example, 3 consecutive detection cycles). This method can effectively reduce the false alarm rate.

[0121] When concept drift is detected, the model is locally updated using incremental fine-tuning techniques. The update strategy is slightly different for different types of models:

[0122] For neural network models (LSTM and GRU):

[0123] First, we freeze all layer parameters except the last two layers. Use the most recent N samples (N is usually set to 1000-5000, dynamically adjusted according to the degree of drift) to build a small dataset. Fine-tune for a small number of epochs (usually 5-10) on this small dataset, and set the learning rate to 1 / 10 of the original learning rate. Use gradient clipping to limit the L2 norm of the gradient to less than 5 to prevent excessive parameter updates.

[0124] For XGBoost model:

[0125] Keep the existing tree structure unchanged and only update the predicted values ​​of leaf nodes. Use the most recent N samples (same as the neural network model) to update leaf nodes. Introduce a learning rate decay strategy, with the initial learning rate set to 0.1 and decayed by 10% after each update.

[0126] For the Prophet model:

[0127] Using a sliding window method, only the most recent M data points are retained (M is usually set to 1 / 4 to 1 / 2 of the size of the original training set). The model parameters, especially the seasonal and trend components, are refitted on this reduced data set. For newly introduced regressors, if their effects are detected to have changed significantly, their coefficients are reestimated.

[0128] Through this incremental learning strategy, the multi-model integration framework can adapt to changes in data distribution in a timely manner and maintain the stability and accuracy of prediction performance. In practical applications, this method significantly improves the model's response speed to emergencies (such as equipment failure, water quality abnormalities, etc.), providing strong support for the intelligent management of drinking water production processes.

[0129] In an optional embodiment,

[0130] The preliminary production plan is optimized by using a multi-objective optimization algorithm with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs. The optimized production scheduling scheme includes:

[0131] Construct a multi-objective optimization model, taking minimization of energy consumption, maximization of capacity utilization and minimization of production costs as optimization objectives, while considering total demand satisfaction and capacity limitations as constraints;

[0132] Initialize the particle swarm, where each particle represents a production scheduling plan;

[0133] An adaptive inertia weight strategy is used to update particle velocity, where the inertia weight changes nonlinearly with the number of iterations to balance global exploration and local development capabilities;

[0134] Using a crowding-based leader selection strategy, the crowding of non-dominated solutions is calculated, and the leader is selected through a tournament selection method to maintain the diversity of solutions;

[0135] The ε-constraint processing mechanism is introduced to compare and update the particles according to the degree of constraint violation, where the ε value gradually decreases during the iteration process.

[0136] An adaptive mutation operator based on Gaussian distribution is applied to enhance the local search capability of the algorithm, and the mutation intensity is adaptively adjusted with the number of iterations;

[0137] Iteratively update the particle position and velocity until the termination condition is reached;

[0138] Generate a Pareto optimal solution set, use parallel coordinates and radar charts to visualize the Pareto frontier, and obtain the optimized production scheduling plan.

[0139] Exemplarily, the present application elaborates on the process of optimizing the preliminary production plan using a multi-objective optimization algorithm, with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs, to obtain an optimized production scheduling plan.

[0140] When building a multi-objective optimization model, three objective functions are defined:

[0141] Energy consumption minimization objective function: Calculate the sum of the power and operating time of all equipment in each time period. Capacity utilization maximization objective function: Calculate the sum of the actual output of all equipment in each time period, divided by the product of the theoretical maximum capacity of all equipment and the total production time. Production cost minimization objective function: Calculate the sum of raw material costs, labor costs and energy costs of all equipment in each time period.

[0142] The constraints include:

[0143] Total demand satisfaction constraint: The sum of the actual output of all equipment in each time period must be greater than or equal to the total demand. Capacity limit constraint: The actual output of each equipment in each time period cannot exceed the product of its theoretical maximum capacity and operating time.

[0144] In order to balance global exploration and local development capabilities, a new adaptive inertia weight strategy is adopted. Specifically, the change of inertia weight with the number of iterations follows a nonlinear function. The function decreases exponentially from the maximum weight value to the minimum weight value, and the rate of change is controlled by a nonlinear adjustment parameter. Through experiments, the optimal parameter settings are determined: the maximum weight is 0.9, the minimum weight is 0.4, and the nonlinear adjustment parameter is 2.5.

[0145] This nonlinear change strategy can maintain a large inertia weight in the early stage of optimization, which is conducive to global exploration; while in the later stage of optimization, the inertia weight gradually decreases, which is conducive to local fine search. Compared with the traditional linear decreasing strategy, our method has an average convergence speed of 15.3% and a final solution quality improvement of 7.8% in 50 independent runs.

[0146] In order to maintain the diversity of Pareto frontier solutions, a leader selection strategy based on crowding is proposed. The specific steps are as follows:

[0147] Compute the crowdedness of each solution in the set of non-dominated solutions:

[0148] For each objective function, the solutions are sorted according to the function value. For the sorted solutions, the normalized distances of adjacent solutions on the objective function are calculated. The sum is calculated to obtain the total congestion degree of each solution.

[0149] The leader is selected using the improved tournament selection method:

[0150] Randomly select k candidate solutions (k=3 works best in the experiment). Compare the dominance relationship and crowding degree of these k solutions. If there is a non-dominance relationship, select the solution with the largest crowding degree as the leader. If there is a dominance relationship, select the solution with the highest Pareto rank as the leader.

[0151] This strategy not only considers the quality of the solution (Pareto level), but also the diversity of the solution (crowding), effectively avoiding the algorithm from falling into the local optimum. In our experiments, compared with the traditional random selection strategy, this method makes the final Pareto frontier more evenly distributed and increases the coverage by about 12%.

[0152] In order to effectively handle the constraints, the ε-constraint processing mechanism is introduced. The core idea of ​​this mechanism is to compare and update according to the degree to which the particles violate the constraints. The specific implementation is as follows:

[0153] Define the measure function of the degree of constraint violation: calculate the sum of the squares of the degree of violation of all constraint functions.

[0154] Set the initial ε value to the maximum value of the constraint violation degree in the initial population.

[0155] In each iteration, the ε value is updated: the ε value decreases nonlinearly with the increase of the iteration number, and the decrease rate is determined by a control parameter.

[0156] During a particle update, if the new position violates the constraint to a greater extent than the current ε value, the update is rejected.

[0157] This mechanism allows a certain degree of constraint violation in the early stage of optimization to increase the search space; and gradually tightens the constraints in the later stage of optimization to ensure the feasibility of the final solution. Our experiments show that compared with the traditional penalty function method, the ε-constraint processing mechanism can find more feasible solutions within the same number of iterations, and the average number of feasible solutions increases by 23.5%.

[0158] In order to enhance the local search capability of the algorithm, an adaptive mutation operator based on Gaussian distribution is designed. The mutation operation is implemented by adding a random value that follows Gaussian distribution to the current position, where the mean of Gaussian distribution is 0 and the standard deviation changes with time.

[0159] The change of standard deviation (i.e., mutation strength) follows a nonlinear function, which makes the mutation strength decrease with the increase of iteration number. Through a large number of experiments, we determined the optimal parameter settings: the initial maximum mutation strength is 0.1 and the adaptive parameter is 1.5.

[0160] This adaptive mutation strategy can maintain a large mutation intensity in the early stage of optimization and increase population diversity; while in the later stage of optimization, the mutation intensity gradually decreases, which is conducive to fine-tuning the solution. Our experimental results show that compared with a fixed mutation intensity, this adaptive strategy can increase the hypervolume index (Hypervolume) of the final Pareto frontier by about 8.7%.

[0161] In order to intuitively display the optimization results, we use parallel coordinate charts and radar charts to visualize the Pareto frontier.

[0162] The steps for drawing a parallel coordinate graph are as follows:

[0163] Normalize the value of each objective function to the interval between 0 and 1. Represent each objective function on parallel coordinate axes. Each Pareto optimal solution is represented by a broken line connecting the coordinate axes.

[0164] The steps for drawing a radar chart are as follows:

[0165] Normalize the objective function value to the range of 0 to 1. Draw equally spaced radial axes with the center of the circle as the origin, and each axis represents an objective function. Each Pareto optimal solution forms a polygon on the radar chart.

[0166] By combining these two visualization methods, decision makers can intuitively compare the trade-offs between different solutions on various objectives. This application develops an interactive visualization tool that allows decision makers to observe the changes in the optimal solution in real time by adjusting the weight of the objective function. This greatly improves the efficiency and transparency of the decision-making process.

[0167] In practical applications, this visualization method helps us discover some interesting insights. For example, we observe a strong negative correlation between minimizing energy consumption and maximizing capacity utilization, while the relationship between minimizing production costs and the other two objectives is relatively complex. These findings provide important references for further optimizing production strategies.

[0168] In an optional embodiment,

[0169] The method further comprises:

[0170] The fuzzy analytic hierarchy process is used to construct a decision support system and calculate the weights of different objectives;

[0171] Use the weighted Chebyshev method to select the final production scheduling solution from the Pareto optimal solution set;

[0172] Output an optimized production scheduling plan that strikes a balance between minimizing energy consumption, maximizing capacity utilization, and minimizing production costs;

[0173] Among them, a fuzzy judgment matrix reflecting the relative importance of multiple optimization objectives is established;

[0174] Calculating a fuzzy comprehensive evaluation value based on the fuzzy judgment matrix; calculating the possibility of a fuzzy number according to the fuzzy comprehensive evaluation value; calculating a weight vector using the possibility of the fuzzy number, wherein each component of the weight vector corresponds to a weight of an optimization target;

[0175] Obtain a Pareto optimal solution set, wherein each solution in the Pareto optimal solution set corresponds to a set of objective function values; for each solution in the Pareto optimal solution set, calculate the difference between each objective function value and its ideal value; divide the difference by the difference between the ideal value and the maximum value of the corresponding objective function to obtain a normalized difference; multiply the normalized difference by the weight of the corresponding optimization target to obtain a weighted normalized difference;

[0176] For each Pareto optimal solution, the maximum value of the weighted normalized difference of all its optimization objectives is taken as the evaluation index of the solution; the Pareto optimal solution with the minimum evaluation index is selected as the final decision result.

[0177] For example, when constructing a fuzzy judgment matrix that reflects the relative importance of multiple optimization objectives, a "dynamic adaptive semantic mapping" technology is proposed. The core of this technology is to establish a dynamic semantic-value mapping relationship, which can automatically adjust the numerical representation of semantic judgment according to the characteristics of the current production environment.

[0178] Specifically, a semantic scale with nine levels is designed, ranging from "extremely unimportant" to "extremely important". Each semantic label no longer corresponds to a fixed fuzzy number, but is associated with a dynamically adjusted fuzzy number interval. The adjustment of this interval is based on the following factors:

[0179] Historical data analysis: The system analyzes past decision results and evaluates the impact of different semantic judgments in actual production. Real-time production status: Considers factors such as current production load and equipment status. Market environment: Includes external factors such as raw material prices and product demand.

[0180] This application develops an intelligent agent based on deep reinforcement learning to dynamically adjust these fuzzy number intervals. Through continuous learning and optimization, the agent can gradually improve the accuracy and adaptability of semantic judgments. For example, during periods of large fluctuations in energy prices, the fuzzy number intervals corresponding to the semantic label "very important" may automatically expand to better capture potential changes in energy consumption target weights. Conversely, during periods of more stable production, these intervals may shrink to improve the accuracy of judgments.

[0181] In practical applications, this dynamic adaptive mechanism significantly improves the quality and consistency of fuzzy judgment. In a complex chemical production system involving five optimization objectives, the consistency rate of the judgment matrix was improved by about 35% compared with the traditional static fuzzy judgment method. At the same time, experts reported that this dynamic adjustment mechanism was more in line with their decision-making intuition.

[0182] In order to more accurately calculate the fuzzy comprehensive evaluation value and target weight, this application proposes a "multi-level fuzzy comprehensive evaluation" method. The core idea of ​​this method is to decompose the complex multi-target evaluation problem into multiple levels, each level adopts a different evaluation strategy, and finally obtain the final comprehensive evaluation result through an integration mechanism.

[0183] Hierarchical decomposition: The optimization objectives are divided into multiple levels according to their nature and mutual relationship. For example, the objectives can be divided into three categories: "economic benefits", "production efficiency" and "environmental impact", and each category can be further subdivided into specific objectives. Intra-layer evaluation: For the objectives within each level, an improved fuzzy comprehensive evaluation method is adopted. An evaluation technology based on interval-type intuitionistic fuzzy sets is introduced to better deal with hesitation and uncertainty in expert judgment. Inter-layer integration: A new "soft hierarchical analysis" method is used to calculate inter-layer weights. This method allows partial overlap in the hierarchy, which is more in line with the complex interactive relationship between objectives in actual production.

[0184] Dynamic weight adjustment: A dynamic adjustment mechanism based on time series analysis is introduced to automatically fine-tune the weights of each level according to historical data and current production status. Global optimization: Finally, the final target weight vector is obtained through a global optimization process based on the particle swarm optimization algorithm. A key innovation of this multi-level approach is the introduction of "target dependency analysis" in the evaluation process. A model based on Bayesian networks is developed to quantify the dependencies between different goals. This enables the system to take into account the synergies and potential conflicts between goals when calculating weights, thereby obtaining a more reasonable and balanced weight distribution.

[0185] In a case study of a large steel plant, this multi-level approach significantly improved the accuracy and interpretability of decision making. Compared with the traditional single-layer fuzzy comprehensive evaluation method, the method of this application improved the accuracy of target weight estimation by about 20%, especially when dealing with highly correlated and potentially conflicting targets. In addition, this hierarchical approach also makes the decision process more transparent, making it easier for decision makers to understand and adjust.

[0186] When using the weighted Chebyshev method to select the final solution from the Pareto optimal solution set, this application proposes a "dynamic reference point" technology to improve the flexibility and adaptability of the method. The core idea of ​​this technology is to dynamically adjust the ideal point and the nadir point so that the calculation of the Chebyshev distance can better reflect the current production needs and constraints.

[0187] The specific steps are as follows:

[0188] Dynamic ideal point calculation: Instead of using a fixed ideal point, a "reachable ideal point" is dynamically calculated based on the current production status and short-term forecast. This point takes into account the current production constraints and possible improvement space in the short term. Adaptive nadir point adjustment: Introducing the concept of an "elastic nadir point" allows the nadir point to float within a certain range. This floating range is dynamically adjusted based on historical data and current production volatility.

[0189] Normalization strategy optimization: A "context-aware" normalization method is proposed. This method not only considers the ideal value and maximum value of the objective function, but also introduces a concept of "expected range" based on production history, so that the normalization process can better reflect the characteristics of actual production.

[0190] Dynamic weight adjustment: A dynamic weight adjustment mechanism based on fuzzy control theory has been developed. This mechanism can fine-tune the weight of each target according to real-time feedback during the production process to cope with short-term production fluctuations and emergencies.

[0191] Multi-criteria decision fusion: In the final selection, this application not only considers the weighted Chebyshev distance, but also introduces other decision criteria, such as robustness index and operational flexibility score. Through a multi-criteria decision model based on fuzzy TOPSIS, these factors are integrated to make the final selection.

[0192] An important feature of this enhanced approach is the introduction of a "post-decision evaluation" mechanism. The system tracks the actual execution effect of each decision and feeds this information back to the decision model, forming a closed-loop optimization process. This not only improves the accuracy of decisions, but also enables the system to continuously learn and self-improve.

[0193] In actual application in a large chemical plant, this enhanced weighted Chebyshev method significantly improved the quality and stability of production scheduling. Compared with traditional methods, the method of this application achieved 7.5%, 6.8% and 8.2% improvements in three key indicators: energy consumption, capacity utilization and production cost, respectively. More importantly, this method has shown strong adaptability and can effectively cope with uncertain factors such as raw material price fluctuations and equipment failures, and control the execution deviation of the production plan at a low level.

[0194] In general, the multi-objective production scheduling optimization method based on improved fuzzy hierarchical analysis and enhanced weighted Chebyshev method proposed in this application is not only innovative in theory, but also shows excellent performance in practical applications. This method provides a new solution for the intelligent scheduling of complex production systems and is expected to play an important role in the fields of industry and intelligent manufacturing.

[0195] In an optional embodiment,

[0196] The optimized production scheduling plan is sent to the control system of each drinking water automation equipment to realize the automated control of the production process, including:

[0197] Obtain optimized production scheduling plan;

[0198] Convert the production scheduling plan into standardized digital instructions, securely encrypt the converted digital instructions, and use X.509 certificate authentication and AES-256 encryption algorithm to ensure data transmission security;

[0199] Transmit the encrypted digital instructions to the distributed control system via the OPC UA protocol;

[0200] Parse the received digital instructions in the distributed control system and dynamically adjust the control strategy, including updating PID control parameters and fuzzy control rules;

[0201] Send the parsed control instructions to the control systems of each drinking water automation equipment;

[0202] In the device control system, intelligent actuators supporting the HART protocol are used to execute control instructions to achieve precise device-level control. Edge computing nodes are deployed, and control algorithms are encapsulated using Docker container technology to achieve real-time control and flexible algorithm upgrades at the device layer.

[0203] For example, this technical solution mainly focuses on how to send the optimized production scheduling plan to the control system of each drinking water automation equipment through a distributed control system (DCS), aiming to realize the intelligent and automated control of the drinking water production process. The solution emphasizes multiple technical links, including data transmission security, parsing and execution of control instructions, and the application of edge computing and container technology. In order to ensure the operability and scalability of the solution, it will be specifically elaborated from the aspects of obtaining scheduling plans, data encryption and transmission, control instruction execution, interaction of intelligent control devices, and real-time control of edge computing.

[0204] First, the system needs to obtain an optimized production scheduling plan. This optimized scheduling plan is usually generated by the Advanced Planning and Scheduling System (APS), which combines multi-dimensional data such as demand forecast, equipment status, and energy consumption of drinking water production. The optimization goals of the scheduling plan may include improving production efficiency, reducing energy consumption, and reducing production downtime.

[0205] For example, the system can predict the water demand for the next three days by analyzing the water consumption in the past week, and generate a specific production scheduling plan based on the equipment status and energy price fluctuations. The plan will include the production lines enabled in each time period, the equipment that needs to be deployed, and the specific production task arrangements.

[0206] After obtaining the optimized production scheduling plan, the next step is to convert it into standardized instructions that the system can recognize. Standardized instructions are usually expressed in a digital format and can be parsed and executed by various modules of the control system. The format of the instructions should comply with industrial standards, such as XML or JSON format, to facilitate subsequent transmission and parsing.

[0207] In order to ensure the security of data transmission, the system needs to encrypt and authenticate the converted instructions. This technical solution uses X.509 certificate authentication and AES-256 encryption algorithm. X.509 certificate is used to verify the identity of both parties in communication and ensure the legitimacy and reliability of instruction transmission; AES-256 encryption algorithm ensures the confidentiality of data during transmission and prevents data from being tampered with or leaked.

[0208] Each control system and device has a unique digital certificate. Before transmission, both parties ensure the legitimacy of the identity by mutually verifying the certificate.

[0209] After the identity verification is passed, the production scheduling instructions will be encrypted into ciphertext and transmitted over the network. The AES-256 encryption algorithm ensures that the data will not be maliciously stolen during the transmission process with its high encryption capability.

[0210] The encrypted digital instructions will be transmitted to the distributed control system (DCS) via the OPC UA (Open Platform Communications Unified Architecture) protocol. The OPC UA protocol is a standardized protocol in the field of industrial automation that supports secure information exchange and cross-platform communication. Its advantages are strong compatibility and built-in security mechanisms, which ensure the integrity and reliability of instructions during transmission.

[0211] For example, instructions are transmitted from the central dispatch system to each sub-site and control node via the OPC UA protocol. During the transmission process, the encryption mechanism of OPC UA further enhances the security of data transmission.

[0212] When the encrypted instructions arrive at the distributed control system, the system first decrypts and parses the instructions. The parsed instructions will be used to dynamically adjust the control strategy, including updating the parameters of the PID (proportional-integral-derivative) controller and the fuzzy control rules.

[0213] PID control parameter update: PID controller is a commonly used controller in industrial processes, which can automatically adjust according to the current error. For example, if the flow setting value of the water pump is 100 cubic meters / hour, and the current flow rate is 90 cubic meters / hour, the PID controller will adjust the pump speed according to the deviation until the flow rate reaches the set value.

[0214] Fuzzy control rule update: Fuzzy control is suitable for complex and nonlinear systems. In the drinking water production process, fuzzy control can optimize decisions based on multiple input variables (such as water quality, flow, pressure, etc.) to ensure the stability and efficiency of the system.

[0215] The parsed control instructions will be sent to the control system of each drinking water automation equipment through DCS. The control system of each device will accurately control the equipment through the received instructions. For example, the system will start or stop a water pump according to the dispatching instructions, adjust the flow or pressure of the water treatment equipment, etc.

[0216] For example, the parsed instruction may be "Start water pump No. 2, set the flow rate to 100 cubic meters per hour, and maintain it for 10 hours." After the instruction is sent to the corresponding device, the device will immediately start to perform the corresponding operation.

[0217] At the device level, smart actuators receive and execute control commands by supporting the HART (Highway Addressable Remote Transducer) protocol, a communication protocol widely used in industrial automation that can transmit digital information without affecting analog signals.

[0218] For example, through the HART protocol, the intelligent actuator can receive instructions issued by the dispatching system and adjust the operating parameters of the equipment according to the instructions, such as the motor speed, valve opening, etc., thereby achieving precise control of the equipment.

[0219] In order to improve the real-time control capability and flexibility of the device layer, the system deploys edge computing nodes at the device layer. Edge computing is to sink computing resources from the cloud to the device side, which can reduce the data transmission bandwidth requirements and reduce the delay of data processing.

[0220] Edge computing nodes use Docker container technology to encapsulate and run control algorithms. Docker is lightweight, portable, and well isolated, ensuring rapid deployment and updating of control algorithms. When necessary, the system can dynamically upgrade control algorithms without affecting the normal operation of the equipment.

[0221] For example, when it is found that the performance of the control algorithm is poor, the system can quickly deploy a new control algorithm through the edge computing node to improve the control accuracy and response speed of the equipment.

[0222] This solution achieves automated control of the drinking water production process through an optimized production scheduling plan. By digitizing, encrypting, transmitting, parsing and executing the scheduling plan, the system can efficiently and accurately control each production link. In addition, by using edge computing and Docker container technology, the system has the ability of real-time control and flexible upgrade, ensuring the efficiency and stability of drinking water production.

[0223] In an optional embodiment,

[0224] Parse the received digital instructions in the distributed control system and dynamically adjust the control strategy, including updating PID control parameters and fuzzy control rules:

[0225] Receiving a digital instruction, wherein the digital instruction is semantically encapsulated in JSON-LD format;

[0226] Parsing the digital instruction using a parser based on a finite state machine, wherein the finite state machine comprises five states: initialization, parsing header, parsing body, verification and execution;

[0227] During the parsing process, the try-catch mechanism is used to capture and classify parsing exceptions;

[0228] According to the analysis results, determine the control strategy that needs to be updated, including PID control parameters and fuzzy control rules;

[0229] For the update of PID control parameters, an adaptive adjustment method based on reinforcement learning is adopted, including:

[0230] Define the state space and take the system error, error change rate and integrated error as state variables;

[0231] Design a reward function that takes into account steady-state error, dynamic response, and control energy consumption;

[0232] Use the Q-learning algorithm to update the Q value and adopt the ε-greedy strategy for action selection;

[0233] Introduce a priority experience replay mechanism to perform importance sampling on samples according to the error size;

[0234] For the update of fuzzy control rules, an adaptive adjustment method based on a combination of genetic algorithm and fuzzy neural network is adopted, including:

[0235] Encode fuzzy rules into chromosomes, including membership function parameters and rule weights;

[0236] Design the fitness function, taking into account the mean square error and the complexity of the rule base;

[0237] Perform genetic operations including roulette wheel and elite-preserving selection, arithmetic crossover, and Gaussian mutation;

[0238] The fuzzy rules optimized by genetic algorithm are used as the initial structure of fuzzy neural network;

[0239] The optimized control parameters and rules are applied to the distributed control system to achieve dynamic adjustment of the control strategy.

[0240] For example, this technical solution focuses on how to parse the received digital instructions in a distributed control system (DCS) and optimize the PID control parameters and fuzzy control rules by dynamically adjusting the control strategy. The solution encapsulates the instructions in JSON-LD format, relies on finite state machine parsing, and combines intelligent technologies such as reinforcement learning and genetic algorithms to achieve adaptive adjustment. The goal of the solution is to improve the flexibility and response efficiency of the control system, especially for control scenarios of nonlinear and time-varying systems.

[0241] In distributed control systems, digital instructions are semantically encapsulated in the JSON-LD (JSON Linked Data) format. JSON-LD is a format with good readability and scalability, which can convert ordinary JSON format into semantically structured data to better support the system's automated parsing and understanding.

[0242] The instruction parsing process relies on a finite state machine (FSM), which contains the following five states:

[0243] Initialization state: The system is ready to start parsing and verify the integrity and basic format of the instructions.

[0244] Parse header: Parse the metadata parts of the instruction (such as `@context` and `@type`) to determine the type of instruction, target device, etc.

[0245] -Analysis body: analyze the core content of the instruction (such as device operation, parameter adjustment, etc.).

[0246] Verification status: Verify the validity of the instruction to ensure that it complies with the system's control strategy and safety requirements.

[0247] Execution status: Based on the analysis results, the instructions are converted into specific control actions and sent to the device.

[0248] During the parsing process, the system uses a try-catch mechanism to capture exceptions that may occur during the parsing process. Common exceptions include incorrect instruction format, missing or illegal parameters, etc. The system will classify and handle them according to the exception type, and implement different recovery mechanisms according to the severity, such as re-requesting instructions or skipping invalid instructions.

[0249] According to the analysis results, the system needs to dynamically adjust the control strategy, mainly including the update of PID control parameters and fuzzy control rules. The core of this part is to realize adaptive adjustment of the control strategy through intelligent algorithms, so as to improve the response performance and steady-state accuracy of the control system.

[0250] For the adjustment of PID control parameters, an adaptive adjustment method based on reinforcement learning is adopted. The specific steps are as follows:

[0251] State space definition: The error of the control system, the rate of change of the error, and the integral value of the error are taken as state variables. These variables can fully reflect the dynamic behavior of the system and help the algorithm determine whether the current control strategy needs to be adjusted.

[0252] For example, in water pump control, the error can be defined as the difference between the actual flow and the set flow, the error change rate reflects the speed of flow regulation, and the integral error is used to measure the long-term accumulation of deviations.

[0253] Reward function design: The design of the reward function needs to comprehensively consider multiple performance indicators of the system, including steady-state error, dynamic response time, and control energy consumption. After each adjustment of the control parameters, the system will calculate the corresponding reward value based on these indicators. The higher the reward value, the better the current control strategy.

[0254] Q-learning algorithm: The system updates the optimal control strategy in each state through the Q-learning algorithm. Q-learning is a model-free reinforcement learning algorithm based on value iteration, which can gradually converge to the optimal control strategy through trial and error learning. At each state update, the system selects an action (i.e. adjusts the PID parameters) and updates the Q value based on the reward value.

[0255] ε-greedy strategy: In order to balance exploration and exploitation, the system adopts the ε-greedy strategy to make decisions when choosing actions. Most of the time, the system will choose the current optimal control strategy (i.e., exploitation), but with a certain probability (determined by ε), the system will choose a random action (i.e., exploration) to find a potentially better control strategy.

[0256] Priority experience replay: In order to improve learning efficiency, the system introduces a priority experience replay mechanism. This mechanism samples the samples according to the size of the error, and gives priority to learning samples with larger errors, thereby accelerating the optimization of the control strategy.

[0257] For example, when the water pump flow error is large, the system will give priority to learning this data to quickly adjust the control parameters and reduce the error.

[0258] For fuzzy control rules, an adaptive adjustment method based on a combination of genetic algorithm and fuzzy neural network is adopted. The specific steps are as follows:

[0259] Fuzzy rule encoding: Encode the fuzzy control rules into chromosomes, which contain the parameters of the membership function and the weights of the rules. Through this encoding method, the system can represent the fuzzy control rules as optimizable individuals, which is convenient for genetic algorithms to operate.

[0260] For example, a fuzzy rule might be “If water pressure is low and water flow is high, then reduce pump speed.” The rule can be encoded by membership function parameters (such as the degree of membership of water pressure and flow) and weights (importance of the rule).

[0261] Fitness function design: The fitness function is used to evaluate the quality of each chromosome. The design of the fitness function needs to comprehensively consider the mean square error of the control system and the complexity of the rule base. The smaller the mean square error, the higher the fitness value; the lower the complexity of the rule base, the higher the fitness value, thus encouraging the system to simplify the rule base while ensuring control accuracy.

[0262] Genetic Operations:

[0263] Selection: Roulette wheel selection and elite retention strategy are used to select the best performing chromosomes to enter the next generation. Roulette wheel selection can ensure that individuals with high fitness have a greater chance of being selected, while the elite retention strategy ensures that the best individuals are not eliminated.

[0264] Crossover: Use arithmetic crossover operation to generate new individuals. The crossover operation can generate better offspring by combining the advantages of parent individuals.

[0265] Mutation: Gaussian mutation operation is used to randomly change the values ​​of certain parameters of the chromosome, thereby increasing the diversity of the population and preventing the algorithm from falling into the local optimum.

[0266] Fuzzy neural network initialization: The fuzzy rules optimized by genetic algorithm will be used as the initial structure of fuzzy neural network. Fuzzy neural network combines the advantages of fuzzy control and neural network, and can achieve more precise control in complex nonlinear systems.

[0267] Apply optimized control parameters and rules: After optimization by genetic algorithm and fuzzy neural network, new control parameters and rules will be applied to the distributed control system to achieve dynamic adjustment of control strategy.

[0268] The control strategy optimized through the above steps will be applied to the distributed control system. The system will control the drinking water automation equipment in real time through the updated PID parameters and fuzzy control rules. In practical applications, the system can continuously optimize the control performance through continuous learning and adjustment to ensure the efficient and stable operation of the drinking water production process.

[0269] For example, suppose the system adjusts the PID control parameters of a water pump through reinforcement learning. The optimized parameters enable the water pump to respond to flow changes more quickly and reduce the adjustment time. The optimization of fuzzy control rules helps the system remain stable under complex working conditions and avoids control hysteresis caused by too many rules.

[0270] This solution realizes intelligent control of distributed control systems through the analysis of digital instructions and the adjustment of dynamic control strategies. The solution uses reinforcement learning to adaptively adjust PID parameters, genetic algorithms and fuzzy neural networks to optimize fuzzy control rules, thereby improving the control accuracy and response speed of the system. Through continuous learning and adjustment, the system can maintain efficient and stable operation in complex industrial environments.

[0271] Figure 2 Schematic diagram of the structure of the drinking water automation equipment production scheduling optimization control system according to the embodiment of the present invention. Figure 2 As shown, the system comprises:

[0272] The first unit is used to obtain real-time production data and historical production data of drinking water automation equipment, where the real-time production data includes raw water quality, equipment operating status and production energy consumption, and the historical production data includes past production orders, equipment maintenance records and seasonal demand fluctuations; based on the acquired real-time production data and historical production data, a machine learning algorithm is used to establish a drinking water production prediction model, which is used to predict production demand and equipment performance within a certain period of time in the future;

[0273] The second unit is used to formulate a preliminary production plan according to the output result of the drinking water production prediction model, combined with the current production orders and inventory conditions, including the start-up time of each production line, production batches and product types; then use a multi-objective optimization algorithm to optimize the preliminary production plan with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs, and obtain an optimized production scheduling plan; the optimized production scheduling plan is sent to the control system of each drinking water automation equipment to realize automated control of the production process;

[0274] The third unit is used to monitor the operating status and production indicators of each equipment in real time during the production process. When an abnormal situation is detected, the preset emergency response process is initiated, and the production scheduling plan is dynamically adjusted according to the actual situation. At the same time, data from the entire production process is collected, including raw material usage, production efficiency, product quality and energy consumption indicators, and these data are fed back to the drinking water production prediction model to continuously optimize and update the parameters of the drinking water production prediction model and improve the prediction accuracy.

[0275] According to a third aspect of the embodiments of the present invention,

[0276] An electronic device is provided, comprising:

[0277] processor;

[0278] a memory for storing processor-executable instructions;

[0279] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0280] According to a fourth aspect of the embodiments of the present invention,

[0281] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0282] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0283] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The production scheduling optimization control method of drinking water automation equipment is characterized by: include: Obtain real-time production data and historical production data of drinking water automation equipment, where real-time production data includes raw water quality, equipment operating status and production energy consumption, and historical production data includes past production orders, equipment maintenance records and seasonal demand fluctuations; based on the acquired real-time production data and historical production data, use machine learning algorithms to establish a drinking water production prediction model, which is used to predict production demand and equipment performance within a certain period of time in the future; According to the output results of the drinking water production prediction model, combined with the current production orders and inventory conditions, a preliminary production plan is formulated, including the start-up time of each production line, production batches and product types; Then, a multi-objective optimization algorithm is used to optimize the preliminary production plan with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs, thereby obtaining an optimized production scheduling plan; The optimized production scheduling plan is sent to the control system of each drinking water automation equipment to realize the automated control of the production process; During the production process, the operating status and production indicators of each device are monitored in real time. When an abnormal situation is detected, the preset emergency response process is initiated, and the production scheduling plan is dynamically adjusted according to the actual situation. At the same time, data from the entire production process is collected, including raw material usage, production efficiency, product quality and energy consumption indicators. These data are fed back to the drinking water production prediction model, and the parameters of the drinking water production prediction model are continuously optimized and updated to improve the prediction accuracy. The optimized production scheduling plan is sent to the control system of each drinking water automation equipment to realize the automated control of the production process, including: Obtain optimized production scheduling plan; Convert the production scheduling plan into standardized digital instructions, securely encrypt the converted digital instructions, and use X.509 certificate authentication and AES-256 encryption algorithm to ensure data transmission security; Transmit the encrypted digital instructions to the distributed control system via the OPC UA protocol; Parse the received digital instructions in the distributed control system and dynamically adjust the control strategy, including updating PID control parameters and fuzzy control rules; Send the parsed control instructions to the control systems of each drinking water automation equipment; In the device control system, intelligent actuators supporting the HART protocol are used to execute control instructions to achieve precise device-level control. Edge computing nodes are deployed, and control algorithms are encapsulated using Docker container technology to achieve real-time control and flexible algorithm upgrades at the device level. Parse the received digital instructions in the distributed control system and dynamically adjust the control strategy, including updating PID control parameters and fuzzy control rules: Receiving a digital instruction, wherein the digital instruction is semantically encapsulated in JSON-LD format; Parsing the digital instruction using a parser based on a finite state machine, wherein the finite state machine comprises five states: initialization, parsing header, parsing body, verification and execution; During the parsing process, the try-catch mechanism is used to capture and classify parsing exceptions; According to the analysis results, determine the control strategy that needs to be updated, including PID control parameters and fuzzy control rules; For the update of PID control parameters, an adaptive adjustment method based on reinforcement learning is adopted, including: Define the state space and take the system error, error change rate and integrated error as state variables; Design a reward function that takes into account steady-state error, dynamic response, and control energy consumption; Use the Q-learning algorithm to update the Q value and adopt the ε-greedy strategy for action selection; Introduce a priority experience replay mechanism to perform importance sampling on samples according to the error size; For the update of fuzzy control rules, an adaptive adjustment method based on a combination of genetic algorithm and fuzzy neural network is adopted, including: Encode fuzzy rules into chromosomes, including membership function parameters and rule weights; Design the fitness function, taking into account the mean square error and the complexity of the rule base; Perform genetic operations including roulette wheel and elite-preserving selection, arithmetic crossover, and Gaussian mutation; The fuzzy rules optimized by genetic algorithm are used as the initial structure of fuzzy neural network; Apply the optimized control parameters and rules to the distributed control system to achieve dynamic adjustment of the control strategy; The preliminary production plan is optimized by using a multi-objective optimization algorithm with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs. The optimized production scheduling scheme includes: Construct a multi-objective optimization model, taking minimization of energy consumption, maximization of capacity utilization and minimization of production costs as optimization objectives, while considering total demand satisfaction and capacity limitations as constraints; Initialize the particle swarm, where each particle represents a production scheduling plan; An adaptive inertia weight strategy is used to update particle velocity, where the inertia weight changes nonlinearly with the number of iterations to balance global exploration and local development capabilities; Using a crowding-based leader selection strategy, the crowding of non-dominated solutions is calculated, and the leader is selected through a tournament selection method to maintain the diversity of solutions; The ε-constraint processing mechanism is introduced to compare and update the particles according to the degree of constraint violation, where the ε value gradually decreases during the iteration process. An adaptive mutation operator based on Gaussian distribution is applied to enhance the local search capability of the algorithm, and the mutation intensity is adaptively adjusted with the number of iterations; Iteratively update the particle position and velocity until the termination condition is reached; Generate a Pareto optimal solution set, use parallel coordinates and radar charts to visualize the Pareto frontier, and obtain an optimized production scheduling plan; The method further comprises: The fuzzy analytic hierarchy process is used to construct a decision support system and calculate the weights of different objectives; Use the weighted Chebyshev method to select the final production scheduling solution from the Pareto optimal solution set; Output an optimized production scheduling plan that strikes a balance between minimizing energy consumption, maximizing capacity utilization, and minimizing production costs; Among them, a fuzzy judgment matrix reflecting the relative importance of multiple optimization objectives is established; Calculating a fuzzy comprehensive evaluation value based on the fuzzy judgment matrix; calculating the possibility of a fuzzy number according to the fuzzy comprehensive evaluation value; calculating a weight vector using the possibility of the fuzzy number, wherein each component of the weight vector corresponds to a weight of an optimization target; Obtain a Pareto optimal solution set, wherein each solution in the Pareto optimal solution set corresponds to a set of objective function values; for each solution in the Pareto optimal solution set, calculate the difference between each objective function value and its ideal value; divide the difference by the difference between the ideal value and the maximum value of the corresponding objective function to obtain a normalized difference; multiply the normalized difference by the weight of the corresponding optimization target to obtain a weighted normalized difference; For each Pareto optimal solution, the maximum value of the weighted normalized difference of all its optimization objectives is taken as the evaluation index of the solution; the Pareto optimal solution with the minimum evaluation index is selected as the final decision result.

2. The method according to claim 1, characterized in that Based on the real-time production data and historical production data obtained, a machine learning algorithm is used to establish a drinking water production prediction model. The model is used to predict production demand and equipment performance in a certain period of time in the future, including: Preprocessing the real-time production data and the historical production data includes decomposing the original time series data into trend terms, seasonal terms and residual terms using an improved seasonality and trend decomposition algorithm, wherein the improved seasonality and trend decomposition algorithm uses an adaptive window size, and the adaptive window size is determined according to the seasonal cycle and the sequence length; Performing feature engineering on the preprocessed data, including extracting time features, generating dynamic lag features, constructing cross features, and extracting Fourier features, wherein the dynamic lag features automatically select the most relevant lag period through autocorrelation analysis; Construct a multi-model integration framework for a drinking water production prediction model, including a base learner layer, an intermediate integration layer, and a final integration layer, wherein the base learner layer includes a long short-term memory network, a gated recurrent unit, a prophet model, and an XGBoost algorithm, the intermediate integration layer uses a dynamic weight allocation mechanism to dynamically adjust the weights according to the performance of each model on a validation set, and the final integration layer uses a stacked generalization method to use an elastic network as a meta-learner; the constructed multi-model integration framework is used to predict production demand and equipment performance within a certain period of time in the future; An incremental learning strategy is adopted for online learning and model updating, including the use of a concept drift detection algorithm to detect changes in data distribution and triggering local model updates when significant changes are detected. The update process uses progressive fine-tuning techniques.

3. A drinking water automation equipment production scheduling optimization control system, used to implement the method described in any one of claims 1-2, characterized in that: include: The first unit is used to obtain real-time production data and historical production data of drinking water automation equipment, where the real-time production data includes raw water quality, equipment operating status and production energy consumption, and the historical production data includes past production orders, equipment maintenance records and seasonal demand fluctuations; based on the acquired real-time production data and historical production data, a machine learning algorithm is used to establish a drinking water production prediction model, which is used to predict production demand and equipment performance within a certain period of time in the future; The second unit is used to formulate a preliminary production plan according to the output result of the drinking water production prediction model, combined with the current production orders and inventory conditions, including the start-up time of each production line, production batches and product types; Then, a multi-objective optimization algorithm is used to optimize the preliminary production plan with the goals of minimizing energy consumption, maximizing capacity utilization and minimizing production costs, thereby obtaining an optimized production scheduling plan; The optimized production scheduling plan is sent to the control system of each drinking water automation equipment to realize the automated control of the production process; The third unit is used to monitor the operating status and production indicators of each equipment in real time during the production process. When an abnormal situation is detected, the preset emergency response process is initiated, and the production scheduling plan is dynamically adjusted according to the actual situation. At the same time, data from the entire production process is collected, including raw material usage, production efficiency, product quality and energy consumption indicators, and these data are fed back to the drinking water production prediction model to continuously optimize and update the parameters of the drinking water production prediction model and improve the prediction accuracy.

4. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 2.

5. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 2 is implemented.

Images