A multi-mode pump set self-learning control system

Abstract

The application provides a multi-mode pump group self-learning control system, which comprises an automatic search control mode and an AI model control mode; the automatic search control mode is used in the stage that a pump house has no data accumulation or insufficient data accumulation, can search the frequency of water pump operation under the premise of meeting safe production, and completes diverse data accumulation; the AI model control mode is used to use a model to deduce an optimal control strategy for optimal control after the training of the model is completed based on the accumulated data. The application can automatically provide optimal combination instructions for pump houses in various scenes, reduces the time, cost and difficulty of manual debugging, can learn and optimize based on the accumulated data, automatically adjusts parameters, optimizes the operating state, and reduces energy consumption.

Description

Technical Field

[0001] This invention relates to the field of pump control technology, specifically to a multi-mode pump self-learning control system. Background Technology

[0002] In water supply, drainage, and sewage treatment, pump sets are among the main energy-consuming devices. Traditional pump set control methods typically rely on fixed parameter settings or manual adjustments. This approach not only struggles to accurately adapt to actual operating conditions, leading to low overall system efficiency, but also causes unnecessary power consumption and equipment wear. With the continuous improvement of industrial automation, traditional pump set control systems can no longer meet the demands of modern industry for high efficiency, energy saving, and intelligence.

[0003] The challenge of quickly recommending a reasonable, stable, and energy-efficient pump set combination for a pumping station through simple configuration, and automatically adjusting parameters, optimizing operating status, and reducing energy consumption without human intervention, has become an urgent problem to be solved. Summary of the Invention

[0004] To address the aforementioned issues, this invention discloses a multi-mode pump set self-learning control system, which can quickly recommend reasonable, stable, and energy-saving pump set combinations for pump stations, and automatically adjust parameters, optimize operating status, and reduce energy consumption without human intervention.

[0005] The specific plan is as follows:

[0006] A multi-mode pump set self-learning control system is characterized by including an automatic search control mode and an AI model control mode. The automatic search control mode is used when there is no data accumulation or insufficient data accumulation in the pump station. It can search for the frequency of pump operation and complete the accumulation of diverse data while ensuring safe production. The AI ​​model control mode is used to perform optimal control by using the model to deduce the optimal control strategy after the model has been trained with accumulated data.

[0007] Furthermore, the steps of the automatic search control mode are as follows:

[0008] S1. Manually set the scheduling flow target and put it into production;

[0009] S2. Based on the optimal strategy and fine-tuning strategy under historical operating conditions, the flow rate of the pump set is quickly brought to the scheduling target.

[0010] S3. Stable operation for a period of time;

[0011] S4. The search operator finds a new control strategy that satisfies the scheduling objective;

[0012] S5. Iterate through steps S3 and S4 until the current production run is completed.

[0013] Furthermore, the purpose of step S2 is to prioritize ensuring the rapid achievement of the target operating conditions after production commences, which is divided into four cases:

[0014] (1) When the current flow rate is within 2% of the target flow rate at the time of commissioning: the scheduling target has been achieved, proceed to the next step;

[0015] (2) The current operating condition has historical operation records: Set the water pump status to the historical optimal strategy;

[0016] (3) The relative difference between the current flow rate and the target flow rate exceeds 5%: calculate the same frequency control strategy based on the physical properties of the water pump;

[0017] (4) The difference between the current flow rate and the target flow rate is within 5%: Randomly select a single pump for a small step adjustment strategy.

[0018] Furthermore, the purpose of step S4 is to continuously try new control schemes for training the environment model. The core idea of ​​the search operator is similar to the depth-first search algorithm of a data structure tree, but with the introduction of a certain degree of randomness. The specific steps are as follows:

[0019] S41. Initialize the control frequency that reaches the scheduling target for the first time after commissioning as the root node. The root node has a search path with twice the number of running pumps. For each path, select a single pump P that is currently running, the search direction D, and the opposite direction D' of D.

[0020] S42. For each search path, each step is executed by changing the frequency of a single pump P by one step in direction D, and another single pump P' is randomly selected to change by one step in direction D'.

[0021] S43. The termination condition for each search path is a pre-configured rule.

[0022] S44. After all paths to a root node have been searched, the best result found during the search process is used as the new root node, and the above search process is repeated.

[0023] Furthermore, the AI ​​model control mode includes environmental model structure, automatic model training and updating, and model usage decisions.

[0024] The environmental model structure includes a water supply flow prediction module achievable under different pump control frequencies and a pump energy consumption prediction module under different pump control frequencies. The inputs of the water supply flow prediction module achievable under different pump control frequencies are variables related to the pump frequency and head, where the head is composed of two variables: suction level and outlet pressure. The inputs of the pump energy consumption prediction module under different pump control frequencies are the pump frequency, head, and flow rate.

[0025] The automatic training and updating of the model includes the following steps:

[0026] S1. Trigger the automatic training program for the model; configure the scheduled task and provide the interface to manually trigger the automatic training program in the platform's interactive interface.

[0027] S2. Construct the training dataset; the dataset consists of data generated within the past year, and outliers are identified and processed. After processing the outliers, the data is saved in revive training format.

[0028] S3. Use Revive to train the environment model; execute the Revive environment model training process, supporting configuration to use default parameters and training the model using hyperparameter search strategies;

[0029] S4. Verify the performance of the environment model. After the environment model is trained, take the time when the model started training as the node, and use the newly generated data as the test set. Perform error analysis on the real data and the model's prediction results. If the test data meets a certain quantity and the mean and variance of its error meet the configured threshold, then the model's performance passes the verification.

[0030] S5. Update the environment model that has passed the test; if the training environment model has passed the test and is better than the current model, then update the environment model for AI control.

[0031] The decision-making process for using the model includes the following steps:

[0032] S1. Based on the target flow rate of the pumping station, select the optimal pump activation combination with the highest single-pump efficiency; the activation of the pumps includes the following four constraints:

[0033] (1) Maintenance and repair constraints: Before production, the interactive interface allows users to select water pumps that are currently under maintenance and repair to ensure that the product can be used normally under maintenance conditions;

[0034] (2) Disable combination constraints: Configuration item, which is a combination of pumps that is excluded during decision-making; multiple combinations can be configured.

[0035] (3) Ensure operational constraints: Configuration item, which is a list of water pumps. When making a decision, the water pumps in the constraint item must be enabled;

[0036] (4) Rotation constraints: Configuration items, including the shortest and longest continuous downtime of a single pump and the permitted range of the longest continuous running time of a single pump;

[0037] Sort these constraints according to their priority, execute and verify them in turn. If a constraint causes a combination that cannot be recommended, skip the constraint and issue a warning.

[0038] S2. Under the optimal pump combination, select the optimal pump control strategy. After training the environment model, first, deduce the flow rate that each frequency combination can achieve under the current head, and select the strategy set S that meets the target flow rate; then, deduce the power consumption per thousand tons of water required for each strategy in S, and select the frequency combination with the lowest power consumption per thousand tons of water as the control strategy.

[0039] The beneficial effects of this invention are as follows:

[0040] 1. It can automatically provide the optimal combination of instructions for pump rooms in various scenarios, reducing the time, cost and difficulty of manual debugging;

[0041] 2. It can learn and optimize based on accumulated data, automatically adjust parameters, optimize operating status, and reduce energy consumption. Attached Figure Description

[0042] Figure 1 This is an overall flowchart of the automatic search control mode in this invention.

[0043] Figure 2 This is a flowchart of the fine-tuning strategy in this invention.

[0044] Figure 3 This is a graph showing the constant frequency characteristic of a centrifugal pump.

[0045] Figure 4 This is a graph showing the frequency conversion characteristic curve of a centrifugal pump.

[0046] Figure 5 This is a combined flow rate-head-efficiency curve for the water pump.

[0047] Figure 6 Data flow diagrams for the traffic prediction model and the energy consumption prediction model. Detailed Implementation

[0048] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0049] This invention provides a multi-mode self-learning control system for pump sets, including an automatic search control mode and an AI model control mode. The automatic search control mode is used when there is no data accumulation or insufficient data accumulation in the pump house. It can search for the operating frequency of the water pump under the premise of meeting safe production and complete the accumulation of diverse data. The AI ​​model control mode is used to perform optimal control by using the model to deduce the optimal control strategy after the model has been trained by accumulating data.

[0050] The automatic search control mode is primarily used when the pump station lacks or has insufficient data accumulation. It automatically searches for the pump's operating frequency and accumulates diverse data while ensuring safe production. The core process in this stage is as follows:

[0051] S1. Manually set the scheduling flow target and put it into production.

[0052] S2. Based on the optimal strategy and fine-tuning strategy under historical operating conditions, the flow rate of the pump set is quickly brought up to the scheduling target.

[0053] S3, running stably for a period of time.

[0054] S4. The search operator finds a new control strategy that satisfies the scheduling objective.

[0055] S5. Iterate through Step 3 and Step 4 until the current production run is completed.

[0056] The purpose of step S2 is to prioritize achieving the target operating condition quickly after production begins, which can be divided into four scenarios:

[0057] (1) When the current flow rate is within 2% of the target flow rate at the time of commissioning: the scheduling target has been achieved and the next step can be carried out.

[0058] (2) The current operating conditions have historical operation records: Set the water pump status to the historical optimal strategy.

[0059] (3) The current flow rate is more than 5% different from the target flow rate: calculate the same frequency control strategy based on the physical properties of the water pump.

[0060] (4) The difference between the current flow rate and the target flow rate is within 5%: Randomly select a single pump for a small step adjustment strategy.

[0061] The purpose of step S4 is to continuously try new control schemes for training the environment model. The core idea of ​​the search operator is similar to the depth-first search algorithm in a data structure tree, but with the introduction of a certain degree of randomness. Specifically:

[0062] S41. Initialize the control frequency that reaches the scheduling target for the first time after commissioning as the root node. The root node has a search path with twice the number of running pumps. For each path, select a single pump P that is currently running, the search direction D (frequency increases or decreases), and the opposite direction D' (frequency decreases or increases).

[0063] S42. For each search path, each step is executed by changing the frequency of a single pump P by one step in direction D, and another running pump P' is randomly selected to change by one step in direction D'.

[0064] S43. The termination condition for each search path is a pre-configured rule, such as the maximum frequency difference between individual pumps, the minimum flow rate of a single pump, the minimum efficiency of a single pump, etc.

[0065] S44. After all paths to a root node have been searched, the best result found during the search process is used as the new root node, and the above search process is repeated.

[0066] In this embodiment, as Figure 1 As shown, the detailed steps of the automatic search control mode are as follows (where the error threshold is 2% and the frequency change step size is 0.5Hz):

[0067] (101) Determine if it is a new production line. If yes, proceed to step (102); otherwise, proceed to step (103).

[0068] (102) Update the production information and initialize the algorithm. Determine whether the following condition is met: |current process - target process| / target flow > error threshold. Otherwise, execute step (104). If the condition is met, query the optimal control strategy for the same working condition and determine whether there is a historical record. If the condition is met, execute step (104). Otherwise, execute step (103).

[0069] (103) Determine whether the following condition is met: |current process - target process| / target flow > error threshold. If yes, execute the fine-tuning strategy and send an adjustment command to the PLC. Otherwise, execute step (104).

[0070] (104) Determine whether the following conditions are met: the current frequency remains unchanged for more than 3 minutes. If so, proceed to step (105); otherwise, keep it unchanged and send an adjustment command to the PLC.

[0071] (105) Record the data generated by the current strategy in stable operation, and determine whether the pump combination currently in operation is reasonable. If so, execute step (106). Otherwise, delete the strategy at the head of the queue from the queue, and determine whether the queue is empty. If so, set: root frequency = the optimal strategy recorded in this production, and rebuild the pump selection queue, and then set: current frequency = root frequency. Otherwise, directly set: current frequency = root frequency, and then execute step (106). Wherein, the root frequency is the frequency with the lowest power consumption per thousand tons of water during the search process under the current production.

[0072] (106) Extract the pump P and direction D at the head of the queue Queue, construct a frequency change strategy (the frequency of pump P changes by one step according to direction D, and another pump that is running is randomly selected and changes by one step in the opposite direction of D), modify the frequency according to the frequency change strategy based on the current frequency, and then send an adjustment command to the PLC.

[0073] The initialization algorithm is as follows:

[0074] 1. Set the "root frequency" to the current running frequency;

[0075] 2. Set the frequency variation step size S;

[0076] 3. Based on the currently running water pumps, construct a pump selection queue with a self-learning strategy, which includes all currently running pumps and the direction of pump frequency change: upward or downward. The queue size is twice the size of the running water pumps, meaning that a pump may have its frequency increase or decrease.

[0077] In this embodiment, as Figure 2 As shown, the detailed steps of the fine-tuning strategy are as follows (where the error threshold is 5% and the frequency change step size is 0.5Hz):

[0078] (201) Determine whether the following condition is met: |current process - target process| / target flow < error threshold. If yes, execute step (202) fine-tuning strategy; otherwise, execute step (203) same frequency control strategy.

[0079] (202) Fine-tune the strategy and determine whether the following condition is met: current process > target process. If so, randomly select a pump and decrease its frequency by one step. If not, select a pump and increase its frequency by one step.

[0080] (203) Same frequency control strategy, calculate in sequence:

[0081] Target total frequency = Target process / Current process * Current total frequency

[0082] Single pump frequency F = Target total frequency / Number of pumps in operation

[0083] And set the running water pump to a single pump frequency F.

[0084] In this embodiment, the AI ​​model control mode, after accumulating data to complete model training in the automatic search control mode, uses the model to deduce the optimal control strategy for optimal control. Its core lies in the structure of the environmental model, automatic training and updating, and the model's usage decisions.

[0085] 1. Model Structure

[0086] For pump control, the controller needs to achieve the target flow rate and save energy. Therefore, the model has two essential modules: 1) predicting the achievable water supply flow rate under different pump control frequencies, and 2) predicting the pump energy consumption under different pump control frequencies.

[0087] To reduce the data informatization requirements of pump stations, the environmental model needs to be simplified as much as possible. From the perspective of the pump's physical properties, the pump's head-flow rate curve and power-flow rate curve are fundamental properties that can be used to determine the pump's operating conditions. The rules governing constant-frequency and variable-frequency operation of pumps are as follows:

[0088] (1) Fixed frequency law: such as Figure 3 The image shows the characteristic curve of a centrifugal pump. Once the pump's frequency (speed) is determined, under healthy operating conditions, it has a unique head (a quantity that uses gravitational potential energy to represent the energy supplied to the water by the pump, measured in meters; it includes velocity head, height head, pressure head, and loss head), efficiency, and power at a given flow rate. Furthermore, generally, there is a one-to-one correspondence between head and flow rate; one efficiency level may correspond to two flow rate conditions, and one power level may correspond to two flow rate conditions.

[0089] (2) Frequency conversion law: such as Figure 4 As shown, there are specific relationships between the frequency (speed) of a water pump and its flow rate, head, efficiency, and power. A water pump has an equivalent operating point at different frequencies. Specifically, if the ratio of two frequencies is n, then the ratio of the flow rates at the two equivalent operating points at the two frequencies is n, the ratio of the heads is n squared, and the ratio of the power is n cubed.

[0090] (3) Pump flow-head-efficiency composite curve: such as Figure 5 As shown, the relationship between flow rate, head, and efficiency is plotted, with flow rate Q, head H, and efficiency η. HQ curves and η-Q curves are plotted in a) and b), respectively. In b), a horizontal line is drawn with a certain efficiency value η1, intersecting the η-Q curves of n1, n2, and n3, resulting in six intersection points. Each intersection point is projected onto the H-Q curve corresponding to the rotational speed, and then these projection points are connected to form a curve, yielding the iso-efficiency curve for η1. Repeating this process yields the iso-efficiency curves for η2, η3, and η4. Similarly, iso-power curves can also be plotted. Figure 5 As shown in Figure a, once the pump's frequency and head are determined, the flow rate and efficiency operating conditions can be deduced from its characteristics. Similarly, its power operating conditions can also be deduced.

[0091] Water supply systems typically use parallel pump sets to meet different flow conditions. The characteristics of the parallel pump set can be obtained by summing up the flow points of each pump corresponding to the same head H value, ignoring pipeline losses.

[0092] Therefore, the specific design of the two prediction models mentioned above is as follows:

[0093] (1) To predict the water supply flow rate achievable under different pump control frequencies, in addition to the frequency, theoretically only the pump head condition needs to be input to deduce the flow rate condition. Therefore, the inputs for this model are the pump frequency and head-related variables. Specifically, since the head loss (energy lost due to friction) is basically proportional to the square of the flow rate, the static head (the height difference between the pump inlet and outlet) will not change. The head is mainly composed of two variables: the suction level and the outlet pressure.

[0094] (2) Predicting pump energy consumption under different pump control frequencies: In addition to frequency, theoretically, only the pump head condition or flow rate condition needs to be input to deduce its energy consumption condition (usually represented by indicators such as power, electricity consumption per thousand tons of water, and electricity consumption for water distribution. These can be calculated and converted between each other). In the input design of this model, redundancy settings are implemented, and both the pump head and flow rate conditions are used as inputs, which makes the model more robust.

[0095] like Figure 6 The actual case of a water source plant deployment shown demonstrates the reliability of the model structure through its actual operation.

[0096] In practical use of self-learning products, under the Revive framework, you only need to configure the dimensions contained in each node of the decision flow graph. For example, the dimension of the pump frequency node pump_freq needs to be the same as the actual number of pumps in the pump room.

[0097] 2. Automatic Model Training and Update

[0098] The automatic training and updating of the model is the most important stage in enabling the AI ​​model control mode to be used. The steps of this process are as follows:

[0099] S1: Trigger the automatic training process for the model.

[0100] Configure a scheduled task, such as triggering an automatic training program every Sunday morning. Additionally, an interface can be provided to allow manual triggering of the automatic training program through the platform's interactive interface.

[0101] S2: Construct the training dataset.

[0102] The dataset primarily uses data generated within the last year. The key challenge in constructing the training set lies in identifying and handling outlier data. For instrument-collected data, outliers typically manifest in three forms: null values, abnormally small values, and abnormally large values. Therefore, configuring appropriate ranges for each data variable can remove most outlier data. After processing the outlier data, the data is saved in a revive training format.

[0103] S3: Use revive to train the environment model.

[0104] After data processing is complete, the revive environment model training process is executed. It supports configuring the model to use default parameters as well as training the model using hyperparameter search strategies.

[0105] S4: Verify the performance of the environment model.

[0106] After the environmental model is trained, it is necessary to evaluate its performance. The time when the model started training is used as a node, and the continuously generated new data is used as the test set. Error analysis is performed between the real data and the model's predictions. If the test data meets a certain quantity (e.g., the test data includes a day's worth of data) and the mean and variance of the errors meet the configured thresholds, then the model's performance passes the evaluation.

[0107] S5: Update the environment model that passed the test.

[0108] If the training environment model passes the test and is better than the current model, then the environment model is updated for AI control.

[0109] 3. Decisions on the use of the model

[0110] The decision-making process for using the model has already been implemented in a real-world case at a water source plant. The specific ideas and steps are as follows:

[0111] S1. Select the optimal pump activation combination based on the target flow rate of the pump room.

[0112] In actual production, the overall efficiency of a pump station is approximately equal to the efficiency of each individual pump multiplied by its flow distribution ratio, and the efficiencies of individual pumps vary. Therefore, when selecting a pump combination, pumps with higher efficiency have higher priority.

[0113] Furthermore, in actual production environments, there are numerous limitations on the selection of water pumps for operation, such as pump unavailability due to maintenance, continuous operating time of a single pump, continuous shutdown time of a single pump, water supply balance in the pump set pipeline, and operational balance of the variable frequency unit. Based on current experience, these constraints can be summarized into the following four categories:

[0114] (1) Maintenance and repair constraints: Before production, the interactive interface allows users to select water pumps that are currently under maintenance and repair to ensure that the product can be used normally under maintenance conditions.

[0115] (2) Disable combination constraints: Configuration item, which is a combination of pumps that is excluded during decision-making. Multiple combinations can be configured.

[0116] (3) Ensure operational constraints: Configuration item, which is a list of water pumps. When making a decision, the water pumps in the constraint item must be enabled.

[0117] (4) Rotation constraints: Configuration items, including the shortest and longest continuous downtime of a single pump and the permitted range of the longest continuous running time of a single pump.

[0118] Sort these constraints according to their priority, execute and verify them in turn. If a constraint causes a combination that cannot be recommended, skip the constraint and issue a warning.

[0119] In summary, after eliminating unselectable pump combinations based on various constraints, this step selects the combination with the highest single-pump efficiency as the optimal pump activation combination.

[0120] S2. Under the optimal pump combination, select the optimal pump control strategy.

[0121] After training the environment model, the flow rate achievable by each frequency combination under the current head is first calculated, and the strategy set S that meets the target flow rate is selected. Then, the power consumption per thousand tons of water required for each strategy in S is calculated, and the frequency combination with the lowest power consumption per thousand tons of water is selected as the control strategy. (Based on the actual deployment of a water source plant, the existing configuration can calculate the result within an average of 0.5 seconds).

[0122] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A multi-mode pump set self-learning control system, characterized in that, It includes an automatic search control mode and an AI model control mode; the automatic search control mode is used when the pump room has no data accumulation or insufficient data accumulation, and searches for the frequency of water pump operation to complete data accumulation under the premise of meeting safe production; the AI ​​model control mode is used to use the model to deduce the optimal control strategy for optimal control after the accumulated data has been used to complete the model training. The steps of the automatic search control mode are as follows: S1. Manually set the scheduling flow target and put it into production; S2. Based on the optimal strategy and fine-tuning strategy under historical operating conditions, the flow rate of the pump set is quickly brought to the scheduling target. S3. Stable operation for a period of time; S4. The search operator finds a new control strategy that satisfies the scheduling objective; S5. Iterate through steps S3 and S4 until the current production run is completed; The specific steps of step S4 are as follows: S41. Initialize the control frequency that reaches the scheduling target for the first time after commissioning as the root node. The root node has a search path with twice the number of running pumps. For each path, select a single pump P that is currently running, the search direction D, and the opposite direction D' of D. S42. For each search path, each step is executed by changing the frequency of a single pump P by one step in direction D, and another single pump P' is randomly selected to change by one step in direction D'. S43. The termination condition for each search path is a pre-configured rule; S44. After all paths to a root node have been searched, the best result found during the search process is taken as the new root node, and the above search process is repeated. The AI ​​model control mode includes the environmental model structure, automatic model training and updating, and model usage decisions.

2. The multi-mode pump group self-learning control system according to claim 1, characterized in that, Step S2 is used to prioritize ensuring the rapid achievement of the target operating condition after production commences, and it is divided into four cases: (1) The difference between the current traffic and the target traffic is within 2%: The scheduling target has been achieved, proceed to the next step; (2) The current operating condition has historical operation records: Set the water pump status to the historical optimal strategy; (3) The relative difference between the current flow rate and the target flow rate exceeds 5%: calculate the same frequency control strategy based on the physical properties of the water pump; (4) The difference between the current flow rate and the target flow rate is within 5%: Randomly select a single pump for a small step adjustment strategy.

3. The multi-mode pump group self-learning control system according to claim 1, characterized in that, The environmental model structure includes a water supply flow prediction module achievable under different pump control frequencies and a pump energy consumption prediction module under different pump control frequencies; wherein... The input to the water supply flow prediction module that can be achieved under different pump control frequencies is a variable related to the pump frequency and head, wherein the head is composed of two variables: suction level and outlet pressure. The inputs to the pump energy consumption prediction module under different pump control frequencies are the pump frequency, head, and flow rate.

4. The multi-mode pump group self-learning control system according to claim 1, characterized in that, The automatic training and updating of the model includes the following steps: S1. Trigger the automatic training program for the model; configure the scheduled task and provide the interface to manually trigger the automatic training program in the platform's interactive interface. S2. Construct the training dataset; the dataset consists of data generated within the past year, and outliers are identified and processed. After processing the outliers, the data is saved in revive training format. S3. Use Revive to train the environment model; execute the Revive environment model training process, supporting configuration to use default parameters and training the model using hyperparameter search strategies; S4. Verify the performance of the environment model; After the environment model is trained, take the time when the model starts training as the node, use the new data generated afterward as the test set, and perform error analysis between the real data and the model's prediction results. If the test data meets a certain quantity and the mean and variance of its error meet the configured threshold, then the model's performance passes the verification. S5. Update the environment model that has passed the test; if the training environment model has passed the test and is better than the current model, then update the environment model for AI control.

5. The multi-mode pump group self-learning control system according to claim 1, characterized in that, The decision-making process for using the model includes the following steps: S1. Based on the target flow rate of the pump room, select the optimal pump combination with the best single pump efficiency; S2. Under the optimal pump combination, select the optimal pump control strategy.

6. The multi-mode pump group self-learning control system according to claim 5, characterized in that, The activation of the water pump is subject to the following four constraints: (1) Maintenance and repair constraints: Before production, check the water pump that is under maintenance and repair in the interactive interface to ensure that the product can be used normally under maintenance conditions; (2) Disable combination constraints: Configuration item, which is a combination of water pumps, and excludes this combination when making a decision; multiple combinations can be configured; (3) Ensure operational constraints: Configuration item, which is a list of water pumps. When making a decision, the water pumps in the constraint item must be enabled; (4) Rotation constraints: Configuration items, including the shortest and longest continuous downtime of a single pump and the permitted range of the longest continuous running time of a single pump; Sort these constraints according to their priority, execute and verify them in turn. If a constraint causes a combination that cannot be recommended, skip the constraint and issue a warning.

7. A multi-mode pump group self-learning control system according to claim 5, characterized in that, After training the environment model, the flow rate that each frequency combination can achieve under the current head is first deduced, and the strategy set S that meets the target flow rate is selected; then the power consumption per thousand tons of water required for each strategy in S is deduced, and the frequency combination with the lowest power consumption per thousand tons of water is selected as the control strategy.

Images