An optimal scheduling method for reducing power cost of sewage treatment

Abstract

The application belongs to the technical field of sewage treatment optimization, and mainly relates to an optimization scheduling method for reducing the power consumption of sewage treatment, which comprises comprehensive data acquisition; data is sent to an edge gateway for preprocessing and to a background, and the background carries out normalization processing on the uploaded data; a long short-term memory network is sent for prediction; a scheduling optimization decision framework is constructed according to the prediction information, a state space and an action space are constructed; a reward function is constructed according to the state space and the action space, a reinforcement learning is introduced to train the scheduling decision, and after the training is completed, the scheduling decision is sent to a sewage treatment plant control system center; after being received by the control system center, an intelligent scheduling strategy model and a reasoning service program are packaged into a Docker image, the packaged reasoning service program is run on the cloud, a state vector formed by the state splicing of the current equipment read by a data acquisition and monitoring control system is sent to the reasoning service program, the reasoning service program outputs a recommended action, and the recommended action is issued to a control unit.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sewage treatment optimization, and mainly relates to an optimization scheduling method for reducing the power cost of sewage treatment. BACKGROUND

[0002] The optimization scheduling method for the power cost of sewage treatment is a method for constructing a prediction model by collecting the operation information of energy-consuming equipment and the effluent quality data, synchronously collecting real-time electricity prices, constructing an operation space and an action space according to the operation state of the equipment as a constraint scheduling decision set of the prediction model, and then introducing reinforcement learning for strategy training, so that the scheduler can adaptively learn the optimal decision strategy.

[0003] In China, the publication number is CN113222666B, an optimization scheduling method for reducing the power cost of sewage treatment. The invention determines the reaction pool participating in sewage treatment and the treatment stage, treatment cycle and electricity price period it is in; loads the power data of the treatment equipment in each treatment stage and the data of the start time point of each electricity price period, the data of the end time point of each electricity price period and the rate data of each electricity price period; determines the decision variables in the scheduling optimization mathematical model; determines the constraint conditions in the scheduling optimization mathematical model in combination with the actual operation conditions of sewage treatment; determines the objective function in the scheduling optimization mathematical model; the constructed scheduling optimization mathematical model is first converted into a program language, and then the software is used for optimal solution.

[0004] In the field of optimization technology, although the invention solves the problem of minimizing the power cost by scheduling optimization to make the aeration stage in the treatment process as much as possible concentrated in the low electricity price period according to the sewage treatment capacity and water quality requirements under the condition of maintaining the existing process and equipment unchanged, but in the case of weather change and seasonal rainfall increase, due to the lack of real-time prediction of weather change, sewage treatment is carried out in the low electricity price period, which is easy to cause sewage treatment aeration overload, the effluent quality cannot meet the standard, aggravate the wear of pipeline, increase the equipment maintenance cost, the low electricity price treatment time is limited, and the sewage treatment has to be arranged in the high electricity price period, resulting in that the actual minimum power consumption is higher than the optimal value outside the rainy season, therefore, an optimization scheduling method for reducing the power cost of sewage treatment is needed, which can predict the weather and realize the optimization of the power consumption of sewage treatment. SUMMARY

[0005] The application provides an optimization scheduling method for reducing the power cost of sewage treatment, which aims to solve the problem that due to the lack of real-time prediction of weather change, sewage treatment is carried out in the low electricity price period, which is easy to cause sewage treatment aeration overload, the effluent quality cannot meet the standard, aggravate the wear of pipeline, increase the equipment maintenance cost, the low electricity price treatment time is limited, and the sewage treatment has to be arranged in the high electricity price period, resulting in that the actual minimum power consumption is higher than the optimal value outside the rainy season.

[0006] In order to solve the above problems, the present application adopts the following technology to realize: an optimal scheduling method for reducing the power cost of sewage treatment: by installing electric energy meters, flow meters, water quality instruments and weather modules, realizing all-around data collection of equipment, process and environment, sending the data to the edge gateway for preprocessing, sending the processed information to the background, normalizing the uploaded data in the background, forming training samples with the normalized feature vectors and historical inflow, sending the training samples to the long short-term memory network for prediction, and outputting load prediction information;

[0007] According to the prediction information, a scheduling optimization decision framework is constructed, the state space and the action space are constructed, and then the state space and the action space are connected according to the state information to form the scheduling optimization decision framework;

[0008] According to the state space and the action space, a reward function is constructed, and a reinforcement learning is introduced to train the scheduling decision, and after the training is completed, the intelligent scheduling strategy model is sent to the control system center of the sewage treatment plant;

[0009] After receiving the intelligent scheduling strategy model, the control system center packs the intelligent scheduling strategy model and the reasoning service program into a Docker image, runs the packaged reasoning service program in the cloud, reads the state vector formed by the current state of the equipment by the data acquisition and monitoring control system, sends it to the reasoning service program, and the reasoning service program outputs the recommended action and issues it to the control unit.

[0010] As a preferred embodiment, the specific steps of realizing all-around data collection of equipment, process and environment are:

[0011] An electric energy meter is installed in the control cabinet of each device to collect real-time power, current and voltage; a flow meter and a water quality online monitor are installed on the pipeline and the reaction tank, and a timing data collection is set; a weather access module is installed in the edge gateway, which is a subsystem for real-time acquisition and provision of weather data.

[0012] As a preferred embodiment, the specific steps of preprocessing the data sent to the edge gateway are:

[0013] The edge gateway performs simple preprocessing on the pulled weather data and the received data, aligns the time stamp according to the system time, eliminates abnormal data records within a reasonable range according to a preset threshold, stores the data in a local queue when the network is temporarily disconnected, and automatically resends the data after the channel is restored; the edge gateway sends the processed data to the monitoring and data acquisition system in the background through the industrial Ethernet by a lightweight publishing message protocol.

[0014] As a preferred embodiment, the specific steps of sending the training samples to the long short-term memory network for prediction are:

[0015] The normalized feature vector and the historical inflow target value are combined into a training sample in the format of (feature vector, historical inflow), and are divided into a training set and a validation set according to time. The divided training sample is written into a time series database in batches for asynchronous reading by a load prediction model. The system sends the prediction to a long short-term memory network by hour, and the long short-term memory network returns the prediction information. According to the returned prediction information, the real-time inflow curve, the current electricity price and the historical electricity price stage distribution, and the key meteorological element trend chart are displayed on the monitoring large screen in the form of charts, and the predicted inflow is monitored in real time.

[0016] As a preferred embodiment, the specific steps of constructing the state space and the action space are: defining the variables of the prediction information, the predicted inflow is , the current electricity price is , the predicted rainfall is , and the current key equipment (aerator, reflux pump, lifting pump) is defined in order. The equipment is D, p is a time index indicating the time corresponding to the current state quantity, and the above variables are combined into a column vector , M is the total number of equipment, is the current state of the Mth equipment, the column vector is stored in a feature database with a time stamp, and the state space is constructed.

[0017] As a preferred embodiment, the specific steps of constructing the state space and the action space further include:

[0018] Defining an executable action set for each equipment , a forward adjustment set of variable frequency gear , and a reverse adjustment set of variable frequency gear ; and combining the actions of all equipment into an action vector , is the variable frequency or switching operation of the uth equipment in the time period p, and a is the executed action; a hierarchical strategy is adopted to determine whether the gear needs to be switched, and then determine whether the specific equipment needs to be switched.

[0019] As a preferred embodiment, the specific steps of constructing the reward function according to the state space and the action space are:

[0020] The reward function has an algorithm formula:

[0021] ,

[0022] is a reward function obtained and timely reward in p time period, used to quantify the action of the current time period, is the electricity cost penalty, is the current electricity price, is the power consumption of the action in the p time period, is the penalty for substandard water quality, is the water quality penalty weight coefficient, indicating the influence of the water quality penalty in the total reward, is the water quality after the action is executed in the p time period, is the threshold for water quality, is the potential overload risk caused by excessive aeration in the rainfall period, is the aeration overload penalty weight coefficient, is the aeration intensity caused by the action in the p time period, EX is the aeration intensity, is the upper limit of the aeration intensity dynamically set based on the rainfall in the p time period.

[0023] As a preferred embodiment, the step of introducing reinforcement learning training scheduling decision to train is:

[0024] From the historical data and real-time data collected by the sewage treatment, simulate the state after the action is executed in each time period, not limited to the update of the equipment action, including water quantity update, equipment update and water quality update, adopt time sequence execution strategy to train, store the data obtained by training to experience pool; Set the storage data capacity FG, when the storage capacity reaches FG, use the proximal policy optimization to iteratively update the policy network parameters, the algorithm formula is:

[0025] ,

[0026] represents the parameter vector of the policy network, is the expected calculation of the experience collected at time v, and the average of the parentheses is calculated on the entire experience pool, is the important sampling ratio, represents the advantage function, is the clipping function, is a hyperparameter.

[0027] As a preferred embodiment, the specific step of the inference service program outputting the recommended action and issuing it to the control unit is:

[0028] After receiving the intelligent scheduling strategy model, the control system center packs the intelligent scheduling strategy model and the inference service program into a Docker image; the inference service program is run in the cloud, and the data acquisition and monitoring control system reads the state vector composed of the predicted inflow , the current electricity price , the predicted rainfall , and the current state of the device; then the state vector is sent to the inference service program, and the recommended action is output, which is sent to the control unit through MQTT.

[0029] The beneficial effects of the present application are:

[0030] 1. By constructing a load prediction model, the rainfall in the future time period can be predicted, providing timely and accurate weather input for scheduling optimization;

[0031] 2. The rainfall is included in the reward function, and the aeration is adjusted during heavy rainfall to avoid the problem of substandard effluent quality caused by excessive aeration during heavy rainfall;

[0032] 3. Real-time weather prediction, more reasonable time period for sewage treatment, and reduced electricity cost for sewage treatment; BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A flow chart of an optimal scheduling method for reducing the electricity cost of sewage treatment. DETAILED DESCRIPTION

[0034] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the following specific embodiments are further described, but the following embodiments are only preferred embodiments of the present application, not all. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. Unless otherwise specified, the experimental methods in the following examples are conventional methods, and the materials, reagents, etc. used in the following examples can be obtained from commercial channels unless otherwise specified.

[0035] Example 1

[0036] As Figure 1 , an optimal scheduling method for reducing the electricity cost of sewage treatment, data acquisition initializes a load prediction model, constructs a scheduling optimization decision framework according to the prediction information, introduces a reinforcement learning algorithm to construct an intelligent scheduling strategy, and implements adaptive optimization by issuing the trained intelligent scheduling strategy. The specific implementation steps are as follows:

[0037] Step 1, data acquisition initializes a load prediction model;

[0038] Step two, build a scheduling optimization decision framework;

[0039] Step three, introduce reinforcement learning algorithm to build intelligent scheduling strategy;

[0040] Step four, issue the model to realize adaptive optimization;

[0041] Among them, in order to realize the data collection of each device, install the electric energy meter in the control cabinet of each device, which is used to collect the electric quantity, current, voltage and so on in real time; In order to realize the data collection of key pipeline and reaction pool, install flowmeter and water quality online monitor on the pipeline and reaction pool, set up timing data collection, such as collecting once every 6 minutes; In order to realize the acquisition of local weather, install meteorological access module in edge gateway, meteorological access module is a subsystem for real-time acquisition and provision of weather data, which pulls the real-time weather data of local area through the public interface API, such as the interface of national meteorological bureau; The electric energy meter and sensor send the acquired data to the edge gateway through industrial Ethernet, the edge gateway carries out simple pretreatment on the acquired weather data and received data, aligns the time stamp according to the system time, eliminates the abnormal data records within the reasonable range according to the preset threshold, the reasonable range refers to the data that does not exceed the threshold and the data that is too high or too low, stores the data in local queue when the network is disconnected temporarily, and automatically retransmits after the channel is restored; The edge gateway sends the processed data to the background monitoring and data acquisition system (SCADA) through industrial Ethernet through the lightweight message queuing telemetry transport (MQTT) protocol, the background stores the high frequency collected electric quantity, flow and water quality index in time sequence database, the time-of-use electricity price is pulled through the electricity price query API provided by local power grid, and the time-of-use electricity price table and weather API pulling results are stored in relational database; The background further pretreats the data in the uploaded data, uses linear interpolation to complete the uploading interruption and vacancy data caused by sensor failure with historical data, aligns the time stamp of all devices and weather, time-of-use electricity price data according to minute, so as to ensure the consistency of time sequence between data, and provides support for subsequent load prediction, scheduling optimization and visual monitoring.

[0042] After further preprocessing of the data, in order to provide more abundant input for the load prediction model, the load prediction model is inputted into the long short-term memory network (LSTM) by the current water inflow and the historical water inflow, and the long short-term memory network is usually used to process and learn time series data; key features are extracted from the completed time series data, and sliding window statistical features including the maximum, minimum and standard deviation of the past average water inflow are constructed; the peak-valley electricity price label is converted into a label vector by the One-Hot encoding method , is an arbitrary time period t, the price of the electricity price P, is the electricity price low valley period, is the flat electricity price period, is the electricity price peak period; meteorological derived features include the cumulative precipitation and temperature change rate in the past period; all continuous quantities are normalized, and the normalization algorithm formula is:

[0043] ,

[0044] R is the normalized feature value, r is the original feature value, is the lower bound of the normalized observation of the original feature in the training, is the upper bound of the normalized observation of the original feature in the training, is the value range of the feature value, represents the lower bound offset of the feature value;

[0045] Continuous quantities refer to those whose values are in a continuous numerical interval, rather than a discrete variable, such as sewage inflow, equipment voltage, online monitoring water quality standard, cumulative precipitation, temperature and humidity. By normalizing the continuous quantities, all feature values are unified to the same interval, avoiding the great influence of some features on the loss function during training, and facilitating subsequent training.

[0046] The normalized feature vector and the historical inflow target value are combined into a training sample in the format of (feature vector, historical inflow), and are divided into a training set and a validation set according to time, and the divided training sample is written into a time series database in batches for asynchronous reading by the load prediction model, which ensures that the load prediction model has good generalization within 24 hours; the system sends the prediction to the long short-term memory network by hour, which is the prediction and analysis of the future 24h inflow, and the long short-term memory network returns the prediction information; according to the returned prediction information, the real-time inflow curve, the current electricity price and the historical electricity price stage distribution, and the key meteorological elements (such as rainfall, temperature) trend chart are displayed on the monitoring large screen in the form of charts, and the predicted inflow is monitored in real time, but if the predicted inflow exceeds the threshold, the early warning module is triggered to send a reminder to the operation and maintenance personnel, prompting possible load increase.

[0047] Step two, after the load prediction model sends the prediction information, a scheduling optimization decision framework is constructed according to the prediction information, which formalizes the prediction information by constructing a state space and an action space to formalize the operation state of the wastewater treatment plant and the executable operation at the moment; in order to provide a complete and quantifiable system snapshot for the scheduling optimization decision framework, the specific steps of constructing the state space are: removing the actions that cause frequent switching of the equipment according to the minimum continuous operation time of the equipment, defining the variables of the prediction information, the predicted inflow is , the current electricity price is , and the predicted rainfall is , and the current key equipment (aerator, reflux pump, lifting pump) is defined in order, the equipment is D, p is a time index indicating the time corresponding to the current state quantity, if the equipment supports variable frequency, the state set is , N is the gear of variable frequency, and if only on-off is supported, the state set is The above variables form a column vector , M is the total number of devices, is the current state of the Mth device, and the column vector is stored in a feature database with a time stamp.

[0048] In order to provide a complete and executable control operation set for the scheduling optimization decision framework, the specific steps of constructing the action space are: defining the executable action set for each device, the forward adjustment set of variable frequency gear is , G is the variable frequency gear of the device, and the reverse adjustment set is ; then all the actions of the devices are combined into an action vector , where is the frequency conversion or switching operation of the u-th device in the time period p, and a is the action performed; in order to make the action space more controllable, a hierarchical strategy is adopted, that is, whether the gear needs to be switched is first judged, and then the specific device is determined.

[0049] The action space and the state space are then connected according to the state information to obtain a clear and accurate device state space, so that the scheduling optimization decision framework can accurately determine the current working condition under the changing flow, electricity price and weather conditions, and formulate a scheduling strategy with the lowest cost and qualified water quality according to the determined working condition. The state space and the action space jointly constitute the scheduling optimization decision framework.

[0050] Step three: after the scheduling optimization decision framework is constructed, in order to make the scheduling strategy self-adaptively and intelligently achieve the optimal balance among the electricity price, the change of water inflow and the water quality standard, reinforcement learning is introduced to train the scheduling optimization decision framework, and the specific steps are as follows: a reward function is constructed according to the state space and the action space constructed in step two, and the algorithm formula is:

[0051] ,

[0052] is the reward function in the p time period, which is used to quantify the advantages and disadvantages of the action in the current time period, is the electricity penalty cost, is the current electricity price, is the power consumption of the action in the p time period, is the penalty for unqualified water quality, is the water quality penalty weight coefficient, which represents the influence of the penalty for unqualified water quality in the total reward, is the water quality after the action is performed in the time period p, is the threshold of qualified water quality, is the potential overload risk caused by excessive aeration in the rainfall period, is the aeration overload penalty weight coefficient, is the aeration intensity caused by the action in the time period p, EX is the aeration intensity, and EX is the aeration intensity, is the upper limit of the aeration intensity dynamically set based on the rainfall in the time period p;

[0053] After the reward function algorithm is constructed, the reinforcement learning is introduced to train the scheduling decision, which mainly includes: based on the historical data and the real-time data collected by the sewage treatment, the action The state after the update is not limited to the update of the execution action of the device, including water quantity update, device update and water quality update, a time sequence execution strategy is adopted for training, and the data obtained by training is stored in an experience pool for caching training samples; a storage data capacity number FG is set, when the storage capacity reaches FG, a proximal policy optimization (PPO) is adopted, the proximal policy optimization algorithm is a policy gradient method considering sample efficiency and stability, which can ensure policy update while considering power cost, water quality and rainfall overload risk, so as to realize optimal scheduling in rainy season and non-rainy season, realize optimization of power cost of sewage treatment, and the proximal policy optimization algorithm formula is:

[0054] ,

[0055] The parameter vector of the policy network is represented as is the expected calculation of the experience collected at time v, and the average of the contents in the brackets is calculated on the entire experience pool, is the important sampling ratio, which represents the change range of the updated policy and the old policy in the same action, represents the advantage function, which describes the benefits of executing the action in the current state, is a clipping function, which is used to limit the sampling ratio to the interval , which is conducive to the stability of policy update and learning process, is a hyperparameter used to control the deviation range of the sampling ratio , is to take the original ratio multiplied by the advantage and the clipped ratio multiplied by the advantage for comparison, if , the smaller ratio after clipping is selected to prevent the ratio from being too large to cause excessive promotion, and if , the larger ratio after clipping is selected to prevent the ratio from being too small to cause excessive punishment.

[0056] During the training process, when facing high-intensity rainfall, the strategy will tend to reduce aeration and delay operation to prevent aeration overload; when the electricity price is in the valley, the strategy will allocate more actions to the high-aeration gear to save power cost; after the training is completed, the intelligent scheduling strategy model is sent to the sewage treatment plant control system center.

[0057] Step four, after receiving the intelligent scheduling strategy model, the control system center packages the intelligent scheduling strategy model and the inference service program into a Docker image. The inference service program is a service module deployed in the cloud to execute the trained reinforcement strategy. The packaged inference service program is run in the cloud, and a control cycle such as six minutes is set. The Supervisory Control and Data Acquisition (SCADA) reads the state vector composed of the predicted inflow , the current electricity price , the predicted rainfall , and the current state of the device; the state vector is sent to the inference service program, and the recommended action is output and sent to the control unit through MQTT; the control unit executes the action after receiving the action and records the execution results such as actual power consumption, device delay, and water quality indicators, and writes the obtained data into the local experience pool; the storage capacity of the experience pool is set to U, and when the storage capacity reaches U, the cyclic coverage strategy is adopted to replace the data of the same time period in chronological order; when the samples in the experience pool continuously exceed the threshold or the water quality is not up to standard, the fine-tuning service is triggered. The fine-tuning service updates the intelligent scheduling strategy model by extracting a small batch of samples from the experience pool and using the nearby policy optimization. In order to better monitor the power consumption, the historical data is trained offline every quarter to extract the seasonal characteristics, which can better discover the impact of rainfall on the cost of wastewater treatment.

[0058] Embodiment 2, based on embodiment 1, an optimal scheduling method for reducing the electricity cost of wastewater treatment, specifically the following scheme:

[0059] In step one, the historical running state, energy consumption, and water quality standards of high-energy-consuming devices in the wastewater treatment plant are collected. A meteorological service adapter is deployed in the cloud to regularly pull the rainfall distribution map and minute-level rainfall forecast from the National Meteorological Administration through an API interface, and to synchronously collect the time-of-use electricity price table and the wastewater inflow. The load and electricity price prediction model is constructed by inputting the long short-term memory network, and the future specific time period inflow and current time period electricity price prediction are output regularly.

[0060] Step two, according to the weather prediction, a rainfall scenario collection is constructed, and the weights of the configuration confidence are divided according to the scenarios, such as moderate rain, light rain, heavy rain, and no rain scenarios. The demand for wastewater inflow is set for each scenario. The future 24 hours are divided into PE time periods. The collected data is normalized, and the minimum objective function is constructed. The weighting between electricity cost, water quality penalty, and device overload penalty is adjusted to achieve the optimal result of minimizing electricity cost and meeting water quality standards.

[0061] Step three, after building the minimum objective function, the system reads the device running state information and the scenario weather forecast of the latest time period in real time, and brings it into the minimum objective function for calculation, and collects the results of the minimum objective function and the actual rainfall in real time. These collected data will be used for updating and iteration of the minimum objective function algorithm; after training, the algorithm is issued to the control center.

[0062] Step four, after receiving the algorithm, the control system center packages the algorithm and the inference service program into a Docker image, runs the packaged inference service program in the cloud, sets the control collection period, such as six minutes, and obtains the current state vector from the data acquisition and monitoring control system; the inference output decision is issued to the control unit.

[0063] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An optimal scheduling method for reducing the electricity cost of sewage treatment, characterized by: By installing electric energy meters, flow meters, water quality instruments and weather modules, collecting equipment, process and environmental data, sending the data to the edge gateway for preprocessing, sending the processed information to the background, normalizing the uploaded data in the background, forming a training sample with the normalized feature vector and historical inflow, sending the training sample to the long short-term memory network for prediction, and outputting the load prediction information; According to the prediction information, a scheduling optimization decision framework is constructed, the state space and action space are constructed, and the state space and action space are connected according to the state information to form the scheduling optimization decision framework; According to the state space and action space, a reward function is constructed, and the algorithm formula of the reward function is: ， is the immediate reward obtained by the reward function at time period p, which quantifies the goodness of the action at the current time period, is the electricity cost penalty, where is the current electricity price, , is the electricity consumption of the action vector at time period p, is the penalty for non-compliance of water quality, is the water quality penalty weight coefficient, which represents the influence of the penalty for non-compliance of water quality in the total reward, is the water quality after performing the action vector at time period p, is the threshold for water quality compliance, is the potential overload risk caused by excessive aeration during the rainfall period, is the aeration overload penalty weight coefficient, is the aeration intensity caused by the action vector at time period p, is the upper limit of aeration intensity which is dynamically set based on the rainfall at time period p. Then introduce reinforcement learning to train the scheduling decision, and after training, send the intelligent scheduling strategy model to the sewage treatment plant control system center; After receiving the intelligent scheduling strategy model, the control system center packs the intelligent scheduling strategy model and the reasoning service program into a Docker image, runs the packaged reasoning service program in the cloud, reads the state vector formed by the current state of the equipment, sends it to the reasoning service program, and the reasoning service program outputs the recommended action and sends it to the control unit.

2. The optimal scheduling method for reducing power consumption of sewage treatment according to claim 1, characterized in that: The specific steps of collecting equipment, process and environmental data are: Install an electric energy meter in the control cabinet of each device to collect real-time power, current and voltage; install flow meters and water quality online monitors on the pipelines and reaction tanks, and set up a timed data collection; install a weather access module in the edge gateway, which is a subsystem for real-time acquisition and provision of weather data.

3. The optimal scheduling method for reducing power consumption of sewage treatment according to claim 1, characterized in that: The specific steps of sending data to the edge gateway for preprocessing are: The edge gateway performs simple preprocessing on the pulled weather data and received data, aligns the time stamp according to the system time, removes abnormal data records within a reasonable range according to the preset threshold, stores the data in the local queue when the network is temporarily disconnected, and automatically resends the data after the channel is restored; The edge gateway sends the processed data to the monitoring and data acquisition system in the background through the industrial Ethernet.

4. The optimal scheduling method for reducing power consumption of sewage treatment according to claim 1, characterized in that: The specific steps of sending the training sample to the long short-term memory network for prediction are: The normalized feature vector and historical inflow target value are combined into a training sample according to the feature vector and historical inflow format, and the divided training sample is written into a time series database by batch; the system sends the long short-term memory network for prediction by hour, and the long short-term memory network returns the prediction information; according to the returned prediction information, it is displayed in the background, the real-time inflow curve, the current electricity price and the historical electricity price stage distribution, and the key meteorological element trend chart are displayed on the monitoring screen in the form of charts, and the predicted inflow is monitored in real time.

5. The optimal scheduling method for reducing power consumption of sewage treatment according to claim 1, characterized in that: The specific steps of constructing the state space and action space are as follows: by defining variables of prediction information, predicting the inflow as , the current electricity price as , and the predicted rainfall as , defining the current key equipment aeration machine, reflux pump and lifting pump in sequence, the equipment being D, p being a time index indicating the time corresponding to the current state quantity, and the above variables forming a column vector , M being the total number of equipment, being the current state of the Mth equipment, storing the column vector in a feature database and attaching a time stamp to construct the state space.

6. The optimal scheduling method for reducing power consumption of sewage treatment according to claim 1, characterized in that: The specific steps of constructing the state space and action space also include: Defining a set of executable actions for each device , a forward adjustment set of variable frequency gear , G is the variable frequency of the gear of the device, and the reverse adjustment set is ; then all the actions of all devices are combined into an action vector , is the variable frequency or switching operation of the u-th device in the time period p; a hierarchical strategy is adopted, that is, whether the gear needs to be switched is determined first, and then whether the specific device needs to be switched.

7. The optimal scheduling method for reducing power consumption of sewage treatment according to claim 1, characterized in that: The steps of introducing reinforcement learning to train the scheduling decision are: Simulate the state after performing the action vector in each time period based on historical data and real-time data collected from sewage treatment , including water quantity update, equipment update and water quality update, adopt time sequence execution strategy for training, and store the data obtained by training to experience pool; A storage data capacity FG is set, when the storage capacity reaches FG, a proximal policy optimization is used to iteratively update the policy network parameters, and a proximal policy optimization algorithm formula is as follows: ， the parameter vector representing the policy network, is the expected computation over the experience collected at time v, averaging over the bracketed terms over the entire experience pool, is the importance sampling ratio, representing the magnitude of change in the updated policy from the old policy for the same action, is the advantage function, indicating the benefit of performing the action in the current state, is the clipping function, to limit the sampling ratio to the interval , is a hyperparameter to control the magnitude of deviation of the sampling ratio .

8. The optimal scheduling method for reducing power consumption of sewage treatment according to claim 1, characterized in that: The inference service program outputs a recommended action, and the specific steps of issuing the recommended action to the control unit are as follows: After receiving the intelligent scheduling strategy model, the control system center packs the intelligent scheduling strategy model and the inference service program into a Docker image; the packed inference service program is run on the cloud, and the data acquisition and monitoring control system reads the state vector composed of the predicted inflow , the current electricity price , the predicted rainfall , and the current state of the device; then the state vector is sent to the inference service program, and the suggested action is output and sent to the control unit through MQTT.

Images