Multi-mode pump set self-learning control system

Abstract

The invention provides a multi-mode pump set self-learning control system. The multi-mode pump set self-learning control system comprises an automatic search control mode and an AI model control mode. The automatic search control mode is used for the stage of no data accumulation or insufficient data accumulation of a pump room, the operation frequency of a water pump can be searched on the premise of meeting safety production, and diversified data accumulation is completed; and the AI model control mode is used for deducing an optimal control strategy by using the model to perform optimal control after data is accumulated to complete model training. According to the invention, the optimal combination instruction can be automatically provided for the pump room in various scenes, and the time, cost and difficulty of manual debugging are reduced; learning optimization can be carried out based on accumulated data, parameters are automatically adjusted, the operation state is optimized, and energy consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of pump group control, and particularly relates to a multi-mode self-learning control system for pump groups. Background Art

[0002] In the fields of water supply, drainage, and sewage treatment, pump groups are one of the main energy-consuming equipment. Traditional pump group control methods usually rely on fixed parameter settings or manual adjustment. This method not only makes it difficult to accurately adapt to the actual working conditions, resulting in low overall system efficiency, but also causes unnecessary power consumption and equipment wear. With the continuous improvement of industrial automation levels, traditional pump group control systems can no longer meet the requirements of modern industries for high efficiency, energy conservation, and intelligence.

[0003] How to quickly recommend a reasonable, stable, and energy-saving pump group combination for the pump house through simple configuration, and be able to automatically adjust parameters, optimize the operating state, and reduce energy consumption without human intervention has become an urgent problem to be solved. Summary of the Invention

[0004] To solve the above problems, the present invention discloses a multi-mode self-learning control system for pump groups, which can quickly recommend a reasonable, stable, and energy-saving pump group combination for the pump house, and automatically adjust parameters, optimize the operating state, and reduce energy consumption without human intervention.

[0005] The specific solutions are as follows: A multi-mode self-learning control system for pump groups, characterized in that it includes an automatic search control mode and an AI model control mode; the automatic search control mode is used for the stage when there is no data accumulation or insufficient data accumulation in the pump house, and can search for the operating frequency of the water pump on the premise of ensuring safe production to complete diverse data accumulation; the AI model control mode is used to use the model to deduce the optimal control strategy for optimal control after the data accumulation is completed for model training.

[0006] Further, 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 historical working condition optimal strategy and the fine-tuning strategy, quickly reach the scheduling target for the flow rate of the pump group; S3. Operate stably for a period of time; S4. Use the search operator to find a new control strategy that meets the scheduling target; S5. Iteratively loop through steps S3 and S4 until the end of this production.

[0007] Further, the purpose of step S2 is to give priority to quickly reaching the target working condition after putting into production, and it is divided into four situations: (1)At the time of production, the difference between the current flow rate and the target flow rate is within 2%: The scheduling target has been achieved, and proceed to the next step; (2)There is a historical operation record for the current working condition: Set the 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 pump; (4)The relative difference between the current flow rate and the target flow rate is within 5%: Randomly select a single pump for a fine-tuning strategy with a small step size.

[0008] Furthermore, the purpose of step S4 is to continuously try new control schemes for the training of the environment model. The core idea of the search operator is similar to the depth search algorithm of a data structure tree, and a certain degree of randomness is introduced. The specific steps are as follows: S41. Initialize the control frequency when production first reaches the scheduling target as the root node. The root node has twice the number of search paths as the number of operating pumps. Each path selects an operating single pump P, a search direction D, and the opposite direction D' of D; S42. For each search path, each execution step is that the frequency of single pump P changes by one step in direction D, and another operating single pump P' is randomly selected to change by one step in direction D'; S43. The end condition for each search path is a pre-configured rule.

[0009] S44. After all paths of a root node have been searched, use the optimal result that appears during the search as the new root node, and repeat the above search process.

[0010] Furthermore, the AI model control mode includes an environment model structure, automatic model training and updating, and model usage decision-making.

[0011] The environment model structure includes a water supply flow prediction module that can be achieved under different pump control frequencies and a pump energy consumption prediction module under different pump control frequencies; among them, the input of the water supply flow prediction module that can be achieved under different pump control frequencies is variables related to the frequency and head of the pump, and the head is composed of two variables, the suction liquid level and the outlet pressure; the input of the pump energy consumption prediction module under different pump control frequencies is the frequency, head, and flow condition of the pump.

[0012] The automatic model training and updating includes the following steps: S1. Trigger the model automatic training program; Configure a timed task and provide an interface to manually trigger the automatic training program on the platform interaction interface; S2. Construct a training dataset; the scope of the dataset is the data generated in the most recent year, and judge and process abnormal data. After completing the processing of abnormal data, save the data in the revive training format; S3. Use the revive training environment model; execute the training process of the revive environment model, support configuration to use default parameters, and train the model using the hyperparameter search strategy; S4. Examine the performance of the environment model.; After the environment model is trained, use the start time of the model training as a node, use the continuously generated new data afterwards as the test set, and perform error analysis on the real data and the prediction results of the model. If the test data meets a certain quantity and the mean and variance of its errors meet the thresholds set in the configuration, the performance of the model passes the inspection; S5. Update the environment model that passes the inspection; if the training environment model passes the inspection and is better than the current model, update the environment model for AI control.

[0013] The usage decision of the said model includes the following steps: S1. Select the optimal water pump enabling combination with the optimal single-pump efficiency according to the target flow of the pump house; the enabling of the water pump includes the following four constraints: (1) Maintenance constraint: Before the interactive interface is put into production, the water pumps under maintenance can be checked to ensure the normal use of the product during maintenance; (2) Disabled combination constraint: A configuration item, which is a water pump combination and is excluded during decision-making; multiple can be configured; (3) Operation guarantee constraint: A configuration item, which is a list of water pumps and the water pumps in the constraint item must be enabled during decision-making; (4) Rotation constraint: A configuration item, which includes the shortest and longest continuous shutdown times of a single pump and the permitted range of the longest continuous operation duration of a single pump; Sort these constraints according to the satisfaction priority, execute and check them in turn. When a certain constraint causes no combination to be recommended, this constraint needs to be skipped and a reminder is required.

[0014] S2. Select the optimal water pump control strategy under the optimal water pump combination. After training the environment model, first deduce the flow rates that can be achieved by each frequency combination at the current head, and screen out the strategy set S that meets the target flow; then deduce the electricity consumption per thousand tons of water required by each strategy in S, and select the frequency combination with the lowest electricity consumption per thousand tons of water as the control strategy.

[0015] The beneficial effects of the present invention are as follows: 1. It can automatically provide the optimal combination instructions for the pump houses in various scenarios, reducing the time, cost and difficulty of manual debugging; 2. It can learn and optimize based on the accumulated data, automatically adjust parameters, optimize the operating state, and reduce energy consumption. Description of the Drawings

[0016] Figure 1 It is the overall flowchart of the automatic search control mode in the present invention.

[0017] Figure 2 It is the flowchart of the fine-tuning strategy in the present invention.

[0018] Figure 3 It is the fixed-frequency characteristic curve of the centrifugal pump.

[0019] Figure 4 It is the variable-frequency characteristic curve of the centrifugal pump.

[0020] Figure 5 It is the comprehensive curve of flow rate - head - efficiency of the water pump.

[0021] Figure 6 Data flow diagram of the flow prediction model and the energy consumption prediction model. Detailed Embodiment

[0022] The following further clarifies the present invention in conjunction with the drawings and the detailed embodiment. It should be understood that the following detailed embodiment is only used to illustrate the present invention and not to limit the scope of the present invention.

[0023] The present invention provides a multi-mode self-learning control system for pump groups, including an automatic search control mode and an AI model control mode; the automatic search control mode is used for the stage when there is no data accumulation or insufficient data accumulation in the pump house, and it can search for the operating frequency of the water pump under the premise of ensuring safe production to complete diverse data accumulation; the AI model control mode is used to use the model to deduce the optimal control strategy for optimal control after the model training is completed based on the accumulated data.

[0024] Among them, the automatic search control mode is mainly used in the stage when there is no data accumulation or insufficient data accumulation in the pump house. It can automatically search for the operating frequency of the water pump under the premise of ensuring safe production to complete diverse data accumulation. The core process of this stage is as follows: S1. Manually set the scheduling flow target and put it into production.

[0025] S2. Based on the optimal strategy of historical working conditions and the fine-tuning strategy, quickly reach the scheduling target for the flow rate of the pump group.

[0026] S3. Operate stably for a period of time.

[0027] S4. Use the search operator to find a new control strategy that meets the scheduling target.

[0028] S5. Iteratively loop through Step 3 and Step 4 until the end of the current production run.

[0029] Among them, the purpose of Step S2 is to prioritize reaching the target operating conditions quickly after production. At this time, it is divided into four cases: (1) When the difference between the current flow rate and the target flow rate at the time of production is within 2%: The scheduling target has been reached, and proceed to the next step.

[0030] (2) There is a historical operation record for the current operating condition: Set the pump state to the historical optimal strategy.

[0031] (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 pump.

[0032] (4) The relative difference between the current flow rate and the target flow rate is within 5%: Randomly select a single pump for a fine-tuning strategy with a small step size.

[0033] The purpose of Step S4 is to continuously try new control schemes for the training of the environmental model. The core idea of the search operator is similar to the depth-first search algorithm of a data structure tree, and a certain degree of randomness is introduced. Specifically: S41. Initialize the control frequency when the production first reaches the scheduling target as the root node. The root node has twice the number of search paths as the number of operating pumps. Each path selects an operating single pump P, a search direction D (frequency increase or decrease), and the opposite direction D' of D (frequency decrease or increase).

[0034] S42. For each search path, each step is to change the frequency of a single pump P by one step size in the direction D, and randomly select another operating pump P' to change the frequency by one step size in the direction D'.

[0035] S43. The end condition for each search path is a pre-configured rule, such as the maximum frequency difference between single pumps, the minimum flow rate of a single pump, the minimum efficiency of a single pump, etc.

[0036] S44. After all paths of a root node have been searched, use the optimal result that appears during the search as the new root node, and repeat the above search process.

[0037] In this embodiment, as Figure 1 shown, the detailed steps of the automatic search control mode are (where the error threshold is 2% and the frequency change step size is 0.5 Hz): (101) Determine whether it is a new production. If so, execute step (102); if not, execute step (103); (102) Update the production start-up information, initialize the algorithm, and determine whether the following condition is met: |Current process - Target process| / Target flow > Error threshold. If not, execute step (104). If so, query the optimal control strategy for the same vertical operating conditions and determine whether there is a historical record. If so, execute step (104). If not, execute step (103); (103) Determine whether the following condition is met: |Current process - Target process| / Target flow > Error threshold. If so, execute the fine-tuning strategy and send an adjustment command to the PLC. If not, execute step (104); (104) Determine whether the following condition is met: The duration of the current frequency remaining unchanged > 3 minutes. If so, execute step (105). If not, keep it unchanged and send an adjustment command to the PLC; (105) Record the data generated during the stable operation of the current strategy and determine whether the currently running pump combination is reasonable. If so, execute step (106). If not, delete the head strategy from the queue Queue and determine whether the queue Queue is empty. If so, set: Root frequency = The optimal strategy recorded in the current production start-up, reconstruct the pump selection queue Queue, and then set: Current frequency = Root frequency. If not, directly set: Current frequency = Root frequency, and then execute step (106); where the root frequency is the frequency with the lowest water and electricity consumption per thousand tons during the search under the current production start-up; (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 size in direction D, and randomly select another running pump to change by one step size in the opposite direction of D), modify the frequency based on the current frequency according to the frequency change strategy, and then send an adjustment command to the PLC.

[0038] Among them, the initialization algorithm is specifically as follows: 1. Set the "root frequency" to the currently running frequency; 2. Set the frequency change step size S; 3. Based on the currently running pumps, construct a pump selection queue Queue for the self-learning strategy, including all currently running pumps and the pump frequency change directions of increase / decrease; the queue size is twice the number of running pumps, that is, the frequency of one pump may increase or decrease.

[0039] In this embodiment, as Figure 2 shown, the detailed steps of the fine-tuning strategy are (where the error threshold is 5% and the frequency change step size is 0.5 Hz): (201) Determine whether the following condition is met: |Current process - Target process| / Target flow < Error threshold. If so, execute step (202) for the fine-tuning strategy. If not, execute step (203) for the same-frequency control strategy; (202)Fine-tuning strategy to determine whether the following condition is met: the current process > the 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. (203)Same-frequency control strategy, calculate successively: Target total frequency = target process / current process * current total frequency Single-pump frequency F = target total frequency / number of operating pumps And set the currently operating pumps to the single-pump frequency F.

[0040] In this embodiment, in the AI model control mode, after accumulating data in the automatic search control mode to complete the training of the model, the model is used to deduce the optimal control strategy for optimal control. The core lies in the structure of the environmental model, automatic training and update, and the decision-making of model use.

[0041] 1. Model structure For pump group control, the controller needs to achieve the target flow rate and save energy. Therefore, the necessary functions of the model have two modules: 1) predicting the water supply flow rate that can be achieved at different pump control frequencies; 2) predicting the pump energy consumption at different pump control frequencies.

[0042] To reduce the requirements for data informatization construction in the pump house, it is necessary to streamline the environmental model as much as possible. In terms of the physical properties of the pump, the head-flow curve and power-flow curve of the pump are the basic properties of the pump and can be used to clarify the operating conditions of the pump. The laws of constant-frequency operation and variable-frequency operation of the pump are as follows: (1)Constant-frequency law: As Figure 3 shown in the characteristic curve specification of a centrifugal pump, after the frequency (speed) of the pump is determined, in the healthy operating state, at a certain flow condition, there is a unique head (Head, the height equivalent representation of all the energy given by the pump to the water using gravitational potential energy, unit: meter. It includes velocity head, height head, pressure head, and loss head), efficiency (Efficiency), and power (Power). In addition, generally speaking, there is a one-to-one correspondence between its head and flow rate. One efficiency may have two corresponding flow conditions, and one power may have two corresponding flow conditions.

[0043] (2)Variable-frequency law: As Figure 4 shown, there is a specific law between the frequency (speed) of the pump and the flow rate, head, efficiency, and power. The pump has equal-efficiency operating points at different frequencies. Specifically, if the ratio of two frequencies is n, then the ratio of the flow rates of the two equivalent operating points at the two frequencies is n, the ratio of the heads is n squared, and the ratio of the powers is n cubed.

[0044] (3)Flow-head-efficiency comprehensive curve of the pump: AsFigure 5 As shown, the variation relationships among flow rate, head, and efficiency are plotted. The flow rate is Q, the head is H, and the efficiency is η. The H-Q curve and the η-Q curve are plotted in a) and b) respectively. In b), a horizontal line is drawn at a certain efficiency value η1, which intersects the η-Q curves of n1, n2, and n3 respectively, with a total of 6 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, thus obtaining the η1 constant-efficiency curve. By repeating the above procedure, the constant-efficiency curves of η2, η3, and η4 are obtained respectively. Similarly, the constant-power curve can also be plotted. As can be seen from Figure 5 Figure a, after the frequency and head of the water pump are determined, the flow rate and efficiency operating conditions can be deduced from its characteristics. Similarly, its power operating condition can also be deduced.

[0045] The water supply system usually uses a parallel pump set to meet the requirements of different flow rate operating conditions. For the parallel law of water pumps, in the case of ignoring pipeline losses, the characteristics of the parallel pump set can be obtained by adding up the flow rate points corresponding to each water pump at the same head H value.

[0046] Therefore, the specific designs of the above two prediction models are as follows: (1) Predict the water supply flow rate that can be achieved at different water pump control frequencies. In addition to the frequency, theoretically, only the head operating condition during its operation needs to be input, and then its flow rate operating condition can be deduced. Therefore, the inputs of this model are variables related to the frequency and head of the water pump. Specifically, since the head loss (the energy lost due to friction) is basically proportional to the square of the flow rate, and the static head (the height difference between the inlet and outlet of the water pump) does not change, the head is mainly composed of two variables: the suction liquid level and the outlet pressure.

[0047] (2) Predict the energy consumption of the water pump at different water pump control frequencies: In addition to the frequency, theoretically, only the head operating condition or the flow rate operating condition during its operation needs to be input, and then its energy consumption operating condition (usually represented by indicators such as power, power consumption per thousand tons of water, and power consumption for water distribution, and they can be calculated and converted with each other) can be deduced. In the input design of this model, a redundant setting is made, and both the head and flow rate operating conditions are used as inputs, which can make the model more robust.

[0048] As Figure 6 shown in the actual case of the deployment of a certain water source plant, the actual operation effect verifies the reliability of the model structure.

[0049] During the actual use of the self-learning product, under the Revive framework, only the dimensions included in each node of the decision flow diagram need to be configured. For example, the dimension of the water pump frequency node pump_freq needs to be the same as the actual number of water pumps in the pump house.

[0050] 2. Automatic Training and Update of the Model The automatic training and updating of the model is the most important stage that the AI model control mode can use. The steps of this process are as follows: S1: Trigger the model automatic training program.

[0051] Configure a timed task, such as triggering the automatic training program at the early morning of every Sunday. In addition, an interface can be provided to manually trigger the automatic training program on the platform interaction interface.

[0052] S2: Build a training dataset.

[0053] The scope of the dataset uses the data generated in the most recent year as the main part. The key points and difficulties in building the training set lie in the judgment and processing of abnormal data. For the data collected by the instrument, its anomalies usually show three forms: null values, abnormally small numerical values, and abnormally large numerical values. Therefore, by configuring the reasonable range of each data variable, most of the abnormal data can be removed. After completing the processing of the abnormal data, save the data in the revive training format.

[0054] S3: Use the revive training environment model.

[0055] After the data processing is completed, execute the training process of the revive environment model. Support configuring to use default parameters and training the model using the hyperparameter search strategy.

[0056] S4: Examine the performance of the environment model.

[0057] After the environment model training is completed, it is necessary to judge the quality of the model. Taking the start time of the model training as a node, use the continuously generated new data afterwards as the test set, and conduct an error analysis on the real data and the prediction results of the model. If the test data meets a certain quantity (for example, the test data contains the data volume of one day) and the mean and variance of its errors meet the thresholds set in the configuration, then the performance of the model passes the inspection.

[0058] S5: Update the environment model that passes the inspection.

[0059] If the training environment model passes the inspection and is better than the current model, then update the environment model for AI control.

[0060] 3. Model usage decision There is already a practical case in a certain water source plant for the model usage decision. Its specific ideas and steps are as follows: S1. Select the optimal combination of pumps to be enabled according to the target flow of the pump house.

[0061] In actual production, the total efficiency of the pump house is approximately equal to the efficiency of each single pump multiplied by its flow distribution ratio, and the efficiencies of single pumps are different. Therefore, during the process of selecting the combination of pumps to be enabled, pumps with higher efficiency have higher priorities.

[0062] In addition, in an actual production environment, there are many restrictions on the selection of pump activation, such as pumps being unavailable for maintenance, the continuous operation time of a single pump, the continuous shutdown time of a single pump, the water transportation balance of the pump group pipeline, the balance of variable frequency unit usage, etc. Based on current experience, the following four constraints are sorted out: (1) Maintenance constraint: Before the interactive interface is put into production, the pumps under maintenance can be selected, ensuring the normal use of the product during maintenance.

[0063] (2) Disabled combination constraint: A configuration item, which is a combination of pumps. This combination is excluded during decision-making. Multiple combinations can be configured.

[0064] (3) Operation guarantee constraint: A configuration item, which is a list of pumps. During decision-making, the pumps in the constraint item must be activated.

[0065] (4) Rotation constraint: A configuration item, including the shortest and longest time that a single pump can be continuously shut down, and the allowable range of the longest continuous operation duration of a single pump.

[0066] Sort these constraints according to the satisfaction priority, execute and check them in sequence. When a certain constraint leads to an inability to recommend a combination, this constraint needs to be skipped and a reminder should be given.

[0067] To sum up, after excluding the ineligible pump combinations according to various constraints in this step, the combination with the optimal single-pump efficiency is selected as the optimal pump activation combination.

[0068] S2. Select the optimal pump control strategy under the optimal pump combination.

[0069] After training the environment model, first deduce the flow rates that can be achieved by each frequency combination under the current head, and screen out the set of strategies S that meet 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. (Based on the actual deployment of a certain water source plant, the existing configuration can calculate the result within an average of 0.5 seconds).

[0070] The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A multi-mode pump group self-learning control system, characterized in that: It includes automatic search control mode and AI model control mode; the automatic search control mode is used in the stage when there is no data accumulation or insufficient data accumulation in the pump room, 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 accumulating data and completing model training.

2. A multi-mode pump group self-learning control system according to claim 1, characterized in that: The steps of the automatic search control mode are: S1. Manually set the dispatch flow target and put it into production; S2, based on the optimal strategy of historical operating conditions and fine-tuning strategy, quickly achieve the dispatching target for the flow of the pump group; S3, stable operation for a period of time; S4, using the search operator to find a new control strategy that meets the scheduling objectives; S5. Iterate and loop through steps S3 and S4 until the current production is completed.

3. A multi-mode pump group self-learning control system according to claim 2, characterized in that: The step S2 is used to prioritize and quickly achieve the target operating condition after production is put into operation, and is divided into four situations: (1) The difference between the current flow and the target flow is within 2%: the scheduling target has been reached, and the next step is carried out; (2) If there is a historical operation record for the current working condition, the pump status is set to the historical optimal strategy; (3) The relative difference between the current flow rate and the target flow rate exceeds 5%: the same frequency control strategy is calculated based on the physical properties of the pump; (4) The relative difference between the current flow rate and the target flow rate is within 5%: a single pump is randomly selected to perform a fine-tuning strategy with small steps.

4. A multi-mode pump group self-learning control system according to claim 2, characterized in that: The specific steps of step S4 are: S41, initialize the control frequency that reaches the scheduling target for the first time after production as the root node, the root node has a search path with twice the number of running pumps, each path selects a single running pump P, a search direction D and the opposite direction D' of D; S42, for each search path, each time a step is executed, the frequency of the single pump P changes by one step in the direction D, and another running single pump P' is randomly selected and changed by one step in the direction D'; S43, the end condition of each search path is a pre-configured rule; S44. After all paths to a root node are searched, the best result found in the search process is used as a new root node, and the above search process is repeated.

5. A multi-mode pump group self-learning control system according to claim 1, characterized in that: The AI ​​model control mode includes environmental model structure, model automatic training and updating, and model usage decision.

6. A multi-mode pump group self-learning control system according to claim 5, characterized in that: The environmental model structure includes a water supply flow prediction module that can be achieved under different water pump control frequencies and a water pump energy consumption prediction module under different water pump control frequencies; wherein, The input of the water supply flow prediction module that can be achieved under different water pump control frequencies is the variables related to the frequency and head of the water pump, and the head is composed of two variables: the water suction level and the outlet pressure; The input of the water pump energy consumption prediction module under different water pump control frequencies is the frequency, head and flow rate conditions of the water pump.

7. A multi-mode pump group self-learning control system according to claim 5, characterized in that: The automatic model training and updating includes the following steps: S1. Trigger the model automatic training program; configure the scheduled task and provide an interface to manually trigger the automatic training program in the platform interactive interface; S2. Build a training data set. The data set is the data generated in the past year. The abnormal data is judged and processed. After the abnormal data is processed, the data is saved in revive training format. S3. Use revive to train the environment model; execute the revive environment model training process, support configuration using default parameters, and use hyperparameter search strategy to train the model; S4. Test the performance of the environment model. After the environment model training is completed, the time when the model training starts is used as the node, and the new data generated continuously afterwards is used as the test set. The error analysis is performed between the real data and the prediction results of the model. If the test data meets a certain number and the mean and variance of the error meet the configured threshold, the performance of the model passes the test. S5. Update the environment model that has passed the inspection; if the training environment model passes the inspection and is better than the current model, update the environment model for AI control.

8. A multi-mode pump group self-learning control system according to claim 5, characterized in that: The decision to use the model includes the following steps: S1. Select the optimal pump activation combination with the best single pump efficiency according to the target flow of the pump room; S2. Under the optimal water pump combination, select the optimal water pump control strategy.

9. A multi-mode pump group self-learning control system according to claim 8, characterized in that: The activation of the pump includes the following four constraints: (1) Maintenance constraints: Before the interactive interface is put into production, you can check the pumps that are under maintenance to ensure that the product can be used normally under maintenance; (2) Disable combination constraint: a configuration item, which is a pump combination. This combination is excluded when making decisions. Multiple combinations can be configured. (3) Guarantee operation constraints: The configuration item is a list of pumps. The pumps in the constraint item must be enabled when making decisions. (4) Rotation constraints: configuration items, including the minimum and maximum time that a single pump can be shut down continuously, and the maximum permitted range of continuous operation time of a single pump; These constraints are sorted according to the priority of satisfaction, executed and checked in sequence. When a constraint makes it impossible to recommend a combination, the constraint needs to be skipped and a reminder is given.

10. A multi-mode pump group self-learning control system according to claim 8, characterized in that: After training the environmental model, we first deduce the flow rate that can be achieved by each frequency combination under the current head, and screen out the strategy set S that meets the target flow rate; then we deduce the water and electricity consumption per thousand tons required for each strategy in S, and select the frequency combination with the lowest water and electricity consumption per thousand tons as the control strategy.

Images