Method, apparatus, medium, and device for electricity optimization of multitask load monitoring and motivation

Abstract

The application relates to a multi-task load monitoring and incentive electricity optimization method, which comprises the following steps: pre-processing a rough-labeled electricity meter data set of an adjustable load to obtain a pre-processed electricity meter data set; dividing the pre-processed electricity meter data set into a training set and a verification set, and training a multi-task load monitoring model on the training set until the loss function value on the verification set converges; deploying the trained multi-task load monitoring model to a user end, analyzing electricity meter data of a power consumption period to obtain an observation quantity, setting a power consumption reward, constructing and training a user optimal power consumption model based on A2C; and putting the trained user optimal power consumption model into application, and when the multi-task load monitoring model detects that the average power size or type of the adjustable load of the user changes, triggering the user optimal power consumption model training process again. The application does not need to rely on an expensive smart home platform or user informed appliance information and use demand, and can be applied to incentive users who can only obtain electricity meter data.

Description

Technical Field

[0001] This application relates to the field of power management technology, and in particular to a power optimization method, apparatus, storage medium and power equipment for multi-task load monitoring and incentive. Background Technology

[0002] With the construction of transparent power distribution networks and the development of smart home technology, the application of smart meters is gradually becoming more widespread. Widely distributed and large-scale adjustable loads on the user side are expected to be incorporated into demand response, fully leveraging their energy-saving benefits and peak-shaving capabilities to contribute to the achievement of "dual carbon" goals. Benefiting from the increase in measurement data, load monitoring technology has become a research hotspot, enabling the perception of appliance-level energy consumption information within a user's home through aggregated meter data. Simultaneously, the implementation of time-of-use pricing has promoted research into optimal electricity consumption methods for users. This involves simultaneously considering external demand response information and the internal electricity needs of users, automatically determining the optimal electricity consumption plan to minimize electricity cost rewards without excessively impacting user comfort.

[0003] After years of development, the limitations of these two technologies have hindered the realization of the potential for adjustable household loads. Firstly, load monitoring technology has a low degree of integration with downstream applications, making it difficult to further utilize the sensed load information. Secondly, user-optimized electricity usage methods lack the ability to sense adjustable loads, relying on users to provide appliance information and usage needs, which cannot address the practical application scenarios of incentive-driven users who can only obtain data from their in-home electricity meters. Therefore, combining load monitoring technology with user-optimized electricity usage technology is crucial for user-side adjustable loads to participate in demand response.

[0004] However, most existing optimal electricity usage methods lack the ability to sense the user's adjustable load. They rely on expensive smart home platforms or depend on users providing information about appliances and usage needs. This approach cannot be applied to incentive-driven users who can only obtain meter data, and it is difficult to automatically detect changes and adjust strategies. Summary of the Invention

[0005] Therefore, it is necessary to provide a power optimization method, device, storage medium, and power equipment that does not rely on expensive smart home platforms or depend on users providing appliance information and usage needs, can be applied to incentive-driven users who can only obtain meter data, and can automatically detect changes and adjust strategies to address the above-mentioned technical problems.

[0006] This invention provides a power consumption optimization method for multi-task load monitoring and incentive, the method comprising the following steps:

[0007] Obtain the coarse-labeled meter dataset of adjustable load, and preprocess the coarse-labeled meter dataset to obtain the preprocessed meter dataset;

[0008] A multi-task load monitoring model is constructed. The preprocessed electricity meter dataset is divided into a training set and a validation set. The multi-task load monitoring model is trained on the training set until the loss function value of the multi-task load monitoring model converges on the validation set.

[0009] The trained multi-task load monitoring model is deployed to the user end, and the meter data of the electricity consumption cycle is analyzed to obtain the types of adjustable loads, average power, and the probability of adjustable load operation in each scheduling period.

[0010] By taking the types of adjustable loads, average power, and external information as observations, and combining the operating probability of adjustable loads in each scheduling period to set electricity consumption rewards, an A2C-based optimal electricity consumption model for users is constructed and trained.

[0011] Once the trained user-optimal electricity consumption model is put into application, when the multi-task load monitoring model detects a change in the average power or type of the user's adjustable load, the user-optimal electricity consumption model training process is re-triggered, the operating probability of the adjustable load in each time period and the type vector of the user's adjustable load are updated, and the electricity consumption reward is modified.

[0012] Furthermore, the adjustable target load includes air conditioners, electric vehicles, and dishwashers; the coarse-labeled meter dataset is a dataset of active power from smart meters with timestamps of load device activation and deactivation; preprocessing of the coarse-labeled meter dataset specifically includes:

[0013] Generate a sequence of 0 and 1 states for each sampling point corresponding to each adjustable load based on the input and output timestamps;

[0014] Calculate the mean μ and standard deviation δ of the active power in the coarsely labeled electricity meter dataset. Subtract the mean from each sampling point and then divide by the variance to complete the data normalization operation.

[0015] The coarsely labeled electricity meter dataset and the switch state 0, 1 sequence are slidably segmented into multiple segments using a sliding window with width w and sliding step size s;

[0016] The average active power p of each adjustable load is calculated within each segment based on the 0, 1 sequence of switch states.

[0017] Furthermore, a multi-task load monitoring model is constructed, specifically including:

[0018] The input fragmented meter data is processed by a one-dimensional convolutional layer to initially extract the intrinsic relationship between adjacent sampling points;

[0019] The long-term operating mode of the load equipment is extracted by the bidirectional gated cyclic network, and then the average active power of various adjustable loads in the segment is mapped by the fully connected layers corresponding to various adjustable loads.

[0020] Multiply the average active power of various adjustable loads by the standard deviation δ of the active power in the coarse-labeled meter data set, and add the mean μ of the active power in the coarse-labeled meter data set to complete the inverse normalization, thus obtaining the multi-task load monitoring model.

[0021] Furthermore, the loss function of the multi-task load monitoring model is:

[0022]

[0023] In the formula, n is the number of types of adjustable loads, and num i Let y be the number of segment samples for the i-th adjustable load. ij Let x be the average active power value of the j-th segment sample in the i-th adjustable load. ij For the normalized active power sequence of the j-th segment sample in the i-th adjustable load, f θ (·) refers to the constructed multi-task load monitoring model.

[0024] Furthermore, the trained multi-task load monitoring model is deployed to the user end, and the meter data of the electricity consumption cycle is analyzed to obtain the types of adjustable loads, average power, and the probability of adjustable load operation in each scheduling period, specifically including:

[0025] The preprocessed meter dataset is input into the multi-task load monitoring model to obtain the average active power of various adjustable loads in each segment;

[0026] The average active power of each segment is calculated to obtain the total average active power p of all adjustable loads in the user's household. m

[0027] Determine the total average active power p m If it is greater than the threshold T p If the user's adjustable load type vector is true, then the corresponding position is set to 1; otherwise, it is set to 0. Here, Pro represents the adjustable load operation probability Pro during the z-th scheduling period of the i-th type of adjustable load. iz The calculation is shown in the following formula:

[0028]

[0029] In the formula p idz Let I be the average active power of the i-th adjustable load during the z-th dispatch period on the d-th day of the electricity consumption cycle; I(·) is a judgment function, which is 1 if the condition in the parentheses is true, and 0 otherwise.

[0030] Furthermore, the external information includes outdoor and indoor temperature information, time-of-use electricity pricing, historical start-stop data for each adjustable load, and user vehicle usage information; the electricity consumption rewards include electricity cost rewards, indoor temperature rewards, user comfort rewards, equipment operation restriction rewards, and electricity demand rewards. The equipment operation restriction rewards include operation restriction rewards for uninterrupted loads. The electricity demand rewards are given after the optimal electricity consumption model provides start-stop suggestions for adjustable loads. If the multi-task load monitoring model detects that the user adopts the corresponding suggestion, then each adjustable load receives a corresponding positive reward; otherwise, it receives a corresponding negative reward.

[0031] The electricity cost reward pri The calculation is shown in the following formula:

[0032] reward pri =-pri next ·p main

[0033] In the formula pri next The time-of-use electricity price at the next scheduling time, p main The active power input to the meter bus in the next scheduling time;

[0034] The indoor temperature reward tem The calculation is shown in the following formula:

[0035]

[0036] In the formula tem next The indoor temperature in Celsius for the next scheduling period;

[0037] The user comfort reward com The calculation is shown in the following formula:

[0038]

[0039] In the formula, act i For the action of the i-th adjustable load in the next scheduling period obtained from the user's optimal power consumption method, act i A value of 1 indicates that the i-th adjustable load is enabled, otherwise it is disabled; z refers to the corresponding index of the next scheduling period; I(·) is a judgment function, which is 1 if the condition in the parentheses is true, otherwise it is 0.

[0040] Furthermore, the user-optimal electricity consumption model includes an actor network and a critic network;

[0041] The optimization objective of the actor network is shown in the following formula:

[0042]

[0043] In the formula, b is the empirical number of parameters involved in gradient calculation in a single parameter update; For the qth experience, the observation during the z-th scheduling period Make an action The reward p represents the observation of the qth experience during the (z+1)th scheduling period; π (·) represents the parameterized model of the actor network, where π represents the corresponding model parameters; V φ (·) represents the parameterized model of the critic network, and φ represents the corresponding model parameters;

[0044] The optimization objective of the critic network is shown in the following formula:

[0045]

[0046] In the formula V φ' (·) represents the critic network before this parameter update, φ' represents the corresponding frozen model parameters, which are treated as constants during the optimization process; γ is the discount factor.

[0047] Another embodiment of the present invention provides a power consumption optimization device for multi-task load monitoring and excitation, the device comprising:

[0048] The data preprocessing module is used to obtain the coarse-labeled meter dataset of adjustable load and preprocess the coarse-labeled meter dataset to obtain the preprocessed meter dataset.

[0049] The model training module is used to construct a multi-task load monitoring model, divide the preprocessed meter dataset into a training set and a validation set, and train the multi-task load monitoring model on the training set until the loss function value of the multi-task load monitoring model on the validation set converges.

[0050] The data analysis module is used to deploy the trained multi-task load monitoring model to the user end and analyze the meter data of the electricity consumption cycle to obtain the types of adjustable loads, average power, and the probability of adjustable load operation in each scheduling period.

[0051] The model building module is used to take the types of adjustable loads, average power, and external information as observations, and to set electricity consumption rewards by combining the operating probability of adjustable loads in each scheduling period, thereby building and training an A2C-based optimal electricity consumption model for users.

[0052] The model optimization module is used to put the trained user-optimal electricity consumption model into application. When the multi-task load monitoring model detects a change in the average power or type of the user's adjustable load, it re-triggers the user-optimal electricity consumption model training process, updates the operating probability of the adjustable load in each time period and the type vector of the user's adjustable load, and modifies the electricity consumption reward.

[0053] Another embodiment of the present invention provides a computer-readable storage medium comprising a stored computer program; wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the power optimization method of multi-task load monitoring and incentive as described above.

[0054] Another embodiment of the present invention provides an electrical device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the power consumption optimization method of multi-task load monitoring and incentive as described above.

[0055] The aforementioned multi-task load monitoring and incentive-based electricity consumption optimization method involves: acquiring a coarse-labeled meter dataset of adjustable loads; preprocessing the coarse-labeled meter dataset to obtain a preprocessed meter dataset; constructing a multi-task load monitoring model by dividing the preprocessed meter dataset into a training set and a validation set; training the multi-task load monitoring model on the training set until the loss function value of the multi-task load monitoring model converges on the validation set; deploying the trained multi-task load monitoring model to the user terminal; and analyzing meter data during electricity consumption cycles to obtain the types and average values ​​of adjustable loads for the user. The system considers power and the probability of adjustable load operation in each scheduling period. It uses the type of adjustable load, average power, and external information as observations, and sets electricity consumption rewards based on the probability of adjustable load operation in each scheduling period. An A2C-based optimal electricity consumption model is constructed and trained. The trained optimal electricity consumption model is then applied. When the multi-task load monitoring model detects a change in the average power or type of adjustable load, the optimal electricity consumption model training process is re-triggered, updating the probability of adjustable load operation in each time period and the type vector of adjustable load, and modifying the electricity consumption rewards. Compared to existing technologies, this invention does not rely on expensive smart home platforms or depend on users providing appliance information and usage needs. It can be applied to incentive-based users who only have access to meter data, and can automatically detect changes and adjust strategies, meeting practical application requirements. Attached Figure Description

[0056] Figure 1 A flowchart illustrating a power consumption optimization method for multi-task load monitoring and excitation provided in an embodiment of the present invention;

[0057] Figure 2This is a schematic diagram of the structure of the multi-task load monitoring model provided in an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of the A2C network structure provided in an embodiment of the present invention;

[0059] Figure 4 This is a structural block diagram of a power consumption optimization device for multi-task load monitoring and excitation provided in an embodiment of the present invention;

[0060] Figure 5 This is a structural diagram of a power equipment provided in an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] It should be noted that the step numbers in this document are only for the convenience of explaining the specific embodiments and are not intended to limit the order in which the steps are executed. The method provided in this embodiment can be executed by a relevant server, and the following description will use a server as the execution subject.

[0063] like Figures 1 to 3 As shown, the power consumption optimization method for multi-task load monitoring and excitation provided in this embodiment of the invention includes steps S11 to S15:

[0064] Step S11: Obtain the coarse-labeled meter dataset of adjustable load, and preprocess the coarse-labeled meter dataset to obtain the preprocessed meter dataset.

[0065] Specifically, this invention can widely collect coarse-labeled meter datasets of adjustable loads based on an intelligent measurement system, and preprocess the coarse-labeled meter datasets. Adjustable loads refer to common electrical appliances with high adjustment potential among household users, characterized by high operating power and variable operating time or power. This invention selects air conditioners, electric vehicles, washing machines, and dishwashers as adjustable loads. The coarse-labeled meter dataset refers to the active power dataset of smart meters with load device connection and disconnection timestamps. In this invention, the sampling interval of the smart meter data is 1 minute. The preprocessing of the coarse-labeled meter dataset involves first generating a sequence of 0 and 1 switching states for each sampling point corresponding to each adjustable load based on the connection and disconnection timestamps; calculating the mean μ and standard deviation δ of the active power in the coarse-labeled meter dataset, subtracting the mean from each sampling point and then dividing by the variance to complete the data normalization operation; then using a sliding window with a width of w and a sliding step size of s to slide and cut the coarse-labeled smart meter data and the 0 and 1 switching state sequences into multiple segments; finally, calculating the average active power p of each adjustable load within each segment based on the 0 and 1 switching state sequences. In this invention, w is taken as 360 and s is taken as 30.

[0066] Step S12: Construct a multi-task load monitoring model, divide the preprocessed meter dataset into a training set and a validation set, and train the multi-task load monitoring model on the training set until the loss function value of the multi-task load monitoring model converges on the validation set.

[0067] As described above, the present invention divides the coarsely labeled electricity meter dataset preprocessed in step S11 into a training set and a validation set, constructs a multi-task load monitoring model based on one-dimensional convolution and bidirectional gated recurrent network, and trains it on the training set until the loss function value on the validation set converges.

[0068] Specifically, the coarse-labeled meter dataset partitioning refers to randomly selecting k% of the sample segments as the training set, and the remaining sample segments as the validation set. In this invention, k is set to 80. The constructed multi-task load monitoring model structure is as follows: Figure 2As shown, the input fragmented meter data is initially extracted through a one-dimensional convolutional layer to extract the intrinsic relationship between adjacent sampling points. A bidirectional gated recurrent network extracts the long-term operating patterns of the load equipment. Then, fully connected layers corresponding to various adjustable loads map the average active power of each adjustable load within the fragment. Finally, each output average is multiplied by the standard deviation δ obtained in step S11 and added to the mean μ obtained in step S11 to complete inverse normalization, obtaining the true predicted value of the average active power of various adjustable loads. Multi-task refers to the fact that this load monitoring model has multiple output layers, simultaneously realizing the regression of the average active power of various adjustable loads. Compared with the common method of establishing an independent model for a single adjustable load, this significantly reduces the overall complexity of the model and improves the generalization ability of the load monitoring model. In this invention, the first one-dimensional convolutional layer has 16 kernels, a kernel size of 1×36, and a stride of 2; the second one-dimensional convolutional layer has 32 kernels, a kernel size of 1×8, and a stride of 2; the input time step of the bidirectional gated recurrent network is 78, and the hidden state dimension is 8; the branch structures in the multi-task output are identical, the number of neurons in the first fully connected layer is 128, and the output dimension of the second fully connected layer is 1; the activation function in this multi-task load monitoring neural network model is the modified linear unit function. The loss function of the multi-task load monitoring model is shown in the following formula:

[0069]

[0070] In the formula, n is the number of types of adjustable loads, and num i Let y be the number of segment samples for the i-th adjustable load. ij Let x be the average active power value of the j-th segment sample in the i-th adjustable load. ij For the normalized active power sequence of the j-th segment sample in the i-th adjustable load, f θ (·) refers to the constructed multi-task load monitoring model. Convergence of the loss function value on the validation set means that the loss function value on the validation set does not decrease for e consecutive rounds; in this invention, e is taken as 20.

[0071] Step S13: Deploy the trained multi-task load monitoring model to the user terminal, and analyze the meter data of the electricity consumption cycle to obtain the types of adjustable loads, average power, and the probability of adjustable load operation in each scheduling period.

[0072] As described above, the multi-task load monitoring model obtained in step S12 is deployed to the user terminal (user's home). After analyzing the electricity meter data for at least one electricity consumption cycle, the load monitoring model obtains the types of adjustable loads, average power, and the probability of adjustable load operation in each scheduling period.

[0073] Specifically, the electricity meter data for one electricity consumption cycle refers to the meter data collected over t consecutive days, including weekends and weekdays. In this invention, t is set to 7. The electricity meter data analysis involves first preprocessing historical data as described in step S11, then inputting it into the multi-task load monitoring model trained in step S12. The output is the average active power of various adjustable loads in each segment. Then, the average active power in each segment is calculated to obtain the total average active power p of various target adjustable loads in the user's household. m Next, determine the total average active power p. m If it is greater than the threshold T p If the user's adjustable load type vector is set to 1, then the corresponding position is set to 0; otherwise, it is set to 0. In this invention, T... p Set to 30W. The operational probability Pro of the adjustable load during the z-th scheduling period for the i-th type of adjustable load. iz The calculation is shown in the following formula:

[0074]

[0075] In the formula p idz Let be the average active power of the i-th adjustable load during the z-th scheduling period on the d-th day of the electricity consumption cycle; I(·) is a judgment function, which is 1 if the condition in parentheses is true, and 0 otherwise. In this invention, the scheduling interval is set to 5 minutes, that is, there are 288 scheduling periods per day.

[0076] Step S14: Using the types of adjustable loads, average power, and external information as observations, and combining the operating probabilities of adjustable loads in each scheduling period to set electricity consumption rewards, a user-optimal electricity consumption model based on A2C is constructed and trained.

[0077] As described above, using the types and average power of adjustable loads obtained in step S13, as well as external information (outdoor and indoor temperature information, time-of-use electricity price, historical start-stop data of each adjustable load, and user vehicle information) as observations, and combining the adjustable load operation probability of each scheduling period obtained in step S13, the optimal electricity consumption reward for users is set, and a user optimal electricity consumption model based on the Advantage Actor-Critic (A2C) network is constructed and trained.

[0078] Specifically, the outdoor and indoor temperature information includes past data. out Outdoor temperature during each scheduling period, future Outdoor temperature forecast for each scheduling period, past... in Indoor temperature bonus for each scheduling period. In this invention, o out , o in All are taken as 12. Time-of-use pricing refers to future electricity prices. priThe time-of-use electricity price sequence for each scheduling period, in this invention o pri Set to 144. Historical start / stop data for each adjustable load refers to past... st The start / stop status sequence of adjustable loads for each decision-making period (0, 1 sequence), in this invention, o st Set to 24. User vehicle usage information refers to the number of time intervals between the next usage time and the current scheduling period, as well as the electric vehicle's state of charge (SOC). Optimal user electricity usage rewards include electricity cost rewards, indoor temperature rewards, user comfort rewards, equipment operation limitation rewards, and electricity demand rewards.

[0079] Electricity cost incentive reward pri The calculation is shown in the following formula:

[0080] reward pri =-pri next ·p main

[0081] In the formula pri next The time-of-use electricity price at the next scheduling time, p main This refers to the active power input to the meter bus in the next scheduling time.

[0082] Indoor temperature reward tem The calculation is shown in the following formula:

[0083]

[0084] In the formula tem next The indoor temperature value in Celsius for the next scheduling period. User comfort reward. com The calculation is shown in the following formula:

[0085]

[0086] In the formula, act i For the action of the i-th adjustable load in the next scheduling period obtained from the user's optimal power consumption method, act i A value of 1 indicates that the i-th adjustable load is enabled, otherwise it is disabled; z refers to the corresponding index of the next scheduling period; I(·) is a judgment function, which is 1 if the condition in the parentheses is true, otherwise it is 0.

[0087] Equipment operation constraint rewards include operation constraint rewards for uninterrupted loads such as washing machines and dishwashers. Invalid action rewards and electric vehicle SOC limit rewards If an uninterruptible load that is currently running is shut down in the next scheduling period, then The reward value is -100, otherwise the reward value is 0; the invalid action reward refers to the reward based on the user's adjustable load type obtained in step 3. If a load that does not exist in the user's home is activated in the next scheduling period, then... Add a -100 bonus; if the next scheduling period coincides with the user's scheduled usage period, the electric vehicle SOC limit bonus will apply. The calculation is shown in the following formula:

[0088]

[0089] In the formula, soc next This represents the state of charge (SOC) value of the electric vehicle battery for the next scheduling period.

[0090] Electricity demand reward need After the optimal power consumption model provides suggestions for the start and stop of adjustable loads, if the multi-task load monitoring model detects that the user adopts the corresponding suggestion, then each adjustable load receives a corresponding positive reward; otherwise, it receives a corresponding negative reward. In this invention, the positive reward values ​​for air conditioners, electric vehicles, washing machines, and dishwashers are 10, 10, 20, and 20, respectively, and the negative reward values ​​are -10, -10, -20, and -20, respectively. The deep reinforcement learning user optimal power consumption model based on the dominant actor-critic network is divided into an actor network and a critic network. The optimization objective of the actor network uses the dominance function as the reward, as shown in the following formula:

[0091]

[0092] In the formula, b is the empirical number of parameters involved in gradient calculation in a single parameter update; For the qth experience, the observation during the z-th scheduling period Make an action The reward p represents the observation of the qth experience during the (z+1)th scheduling period; π (·) represents the parameterized model of the actor network, where π represents the corresponding model parameters; V φ (·) represents the parameterized model of the critic network, and φ represents the corresponding model parameters. The optimization objective of the critic network is shown in the following equation:

[0093]

[0094] In the formula V φ' (·) represents the critic network before the parameter update, φ' represents the corresponding frozen model parameters, which are constants during optimization; γ is the discount factor, which is set to 0.9 in this invention. The output action space of this optimal electricity consumption model is to turn on or off at the next scheduling time for multiple target adjustable loads. In this invention, four types of adjustable loads are selected, including air conditioners, electric vehicles, washing machines, and dishwashers. The dominant actor-critic network structure is as follows: Figure 3As shown, a multi-task architecture with a shared feature extraction layer is adopted, followed by corresponding task modules that output actions and values. The model consists of fully connected layers, with the feature extraction layer having 278, 1024, and 512 neurons respectively; the two task modules have the same structure, with 128 and 4 neurons in their respective layers. Except for the output layer, which uses the Sigmoid activation function, all other layers use the modified linear unit function. If the probability value of the action output is greater than 0.5, it is set to 1, meaning the adjustable load will be turned on or continue to run in the next scheduling time; otherwise, it is set to 0, meaning the adjustable load will be cut off or remain off in the next scheduling time.

[0095] Step S15: The trained user-optimal electricity consumption model is put into application. When the multi-task load monitoring model detects a change in the average power or type of the user's adjustable load, the user-optimal electricity consumption model training process is re-triggered, the operating probability of the adjustable load in each time period and the type vector of the user's adjustable load are updated, and the electricity consumption reward is modified.

[0096] Specifically, the actor network trained in step S14 is put into application. When the multi-task load monitoring model obtained in step S12 detects a change in the average power or type of the user's adjustable load, the A2C training process in step S14 is re-triggered, updating the operating probabilities of the adjustable loads at each time period and the type vector of the user's adjustable loads calculated in step S13, and modifying the equipment operation restriction reward in step S14. In other words, this invention combines load detection technology with a user-optimal electricity consumption method based on deep reinforcement learning in a cascaded manner, fully leveraging the intelligent sensing function of load monitoring technology and the decision-making function of deep reinforcement learning algorithms in uncertain scenarios. A multi-task load detection model is designed for various adjustable loads, sharing a feature extraction layer, inputting fragmented meter data, and simultaneously outputting the average active power value of various target adjustable loads. In practical applications, the load monitoring model is used to monitor changes in the user-side adjustable load in real time, triggering updates to the user-optimal electricity consumption model and other variables, and adapting to the user's adjustable load in a timely manner.

[0097] Understandably, this invention combines load monitoring technology with a user-optimal power consumption method based on deep reinforcement learning. It enables automated optimal power consumption decisions for incentivized users based on accurate perception of internal adjustable loads. A multi-task load monitoring model for adjustable loads is proposed, sharing feature extraction layers for various adjustable loads, which improves the generalization of the load monitoring model while reducing its complexity. Addressing the uncertainties of user power consumption and the external environment, an optimal power consumption method for incentivized users is constructed based on A2C, fully leveraging energy-saving potential while ensuring user comfort rewards are almost unaffected. A2C uses a dominance function as the reward term, measuring the difference between the subsequent reward and the average reward for selecting this action, thus improving training performance. The established multi-task load monitoring model can track changes in user-adjustable loads in real time and adaptively adjust the optimal power consumption strategy to cope with changes in user-side equipment.

[0098] The aforementioned multi-task load monitoring and incentive-based electricity consumption optimization method involves: acquiring a coarse-labeled meter dataset of adjustable loads; preprocessing the coarse-labeled meter dataset to obtain a preprocessed meter dataset; constructing a multi-task load monitoring model by dividing the preprocessed meter dataset into a training set and a validation set; training the multi-task load monitoring model on the training set until the loss function value of the multi-task load monitoring model converges on the validation set; deploying the trained multi-task load monitoring model to the user terminal; and analyzing meter data during electricity consumption cycles to obtain the types and average values ​​of adjustable loads for the user. The system considers power and the probability of adjustable load operation in each scheduling period. It uses the type of adjustable load, average power, and external information as observations, and sets electricity consumption rewards based on the probability of adjustable load operation in each scheduling period. An A2C-based optimal electricity consumption model is constructed and trained. The trained optimal electricity consumption model is then applied. When the multi-task load monitoring model detects a change in the average power or type of adjustable load, the optimal electricity consumption model training process is re-triggered, updating the probability of adjustable load operation in each time period and the type vector of adjustable load, and modifying the electricity consumption rewards. Compared to existing technologies, this invention does not rely on expensive smart home platforms or depend on users providing appliance information and usage needs. It can be applied to incentive-based users who only have access to meter data, and can automatically detect changes and adjust strategies, meeting practical application requirements.

[0099] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0100] Please see Figure 4 The present invention also provides a power consumption optimization device for multi-task load monitoring and excitation, the device comprising:

[0101] The data preprocessing module 21 is used to obtain the coarse-labeled meter dataset of adjustable load and preprocess the coarse-labeled meter dataset to obtain the preprocessed meter dataset.

[0102] The adjustable target load includes air conditioners, electric vehicles, and dishwashers; the coarse-labeled meter dataset is a dataset of active power from smart meters with timestamps of load device activation and deactivation; preprocessing of the coarse-labeled meter dataset specifically includes:

[0103] Generate a sequence of 0 and 1 states for each sampling point corresponding to each adjustable load based on the input and output timestamps;

[0104] Calculate the mean μ and standard deviation δ of the active power in the coarsely labeled electricity meter dataset. Subtract the mean from each sampling point and then divide by the variance to complete the data normalization operation.

[0105] The coarsely labeled electricity meter dataset and the switch state 0, 1 sequence are slidably segmented into multiple segments using a sliding window with width w and sliding step size s;

[0106] The average active power p of each adjustable load is calculated within each segment based on the 0, 1 sequence of switch states.

[0107] The model training module 22 is used to construct a multi-task load monitoring model, divide the preprocessed meter dataset into a training set and a validation set, and train the multi-task load monitoring model on the training set until the loss function value of the multi-task load monitoring model converges on the validation set.

[0108] Furthermore, a multi-task load monitoring model is constructed, specifically including:

[0109] The input fragmented meter data is processed by a one-dimensional convolutional layer to initially extract the intrinsic relationship between adjacent sampling points;

[0110] The long-term operating mode of the load equipment is extracted by the bidirectional gated cyclic network, and then the average active power of various adjustable loads in the segment is mapped by the fully connected layers corresponding to various adjustable loads.

[0111] Multiply the average active power of various adjustable loads by the standard deviation δ of the active power in the coarse-labeled meter data set, and add the mean μ of the active power in the coarse-labeled meter data set to complete the inverse normalization, thus obtaining the multi-task load monitoring model.

[0112] The loss function of the multi-task load monitoring model is:

[0113]

[0114] In the formula, n is the number of types of adjustable loads, and num i Let y be the number of segment samples for the i-th adjustable load. ij Let x be the average active power value of the j-th segment sample in the i-th adjustable load. ij For the normalized active power sequence of the j-th segment sample in the i-th adjustable load, f θ (·) refers to the constructed multi-task load monitoring model.

[0115] The data analysis module 23 is used to deploy the trained multi-task load monitoring model to the user end and analyze the meter data of the electricity consumption cycle to obtain the types of adjustable loads, average power, and the probability of adjustable load operation in each scheduling period.

[0116] The trained multi-task load monitoring model is deployed to the user end, and meter data during the electricity consumption cycle is analyzed to obtain the types of adjustable loads, average power, and the probability of adjustable load operation during each scheduling period. Specifically, this includes:

[0117] The preprocessed meter dataset is input into the multi-task load monitoring model to obtain the average active power of various adjustable loads in each segment;

[0118] The average active power of each segment is calculated to obtain the total average active power p of all adjustable loads in the user's household. m

[0119] Determine the total average active power p m If it is greater than the threshold T p If the user's adjustable load type vector is true, then the corresponding position is set to 1; otherwise, it is set to 0. Here, Pro represents the adjustable load operation probability Pro during the z-th scheduling period of the i-th type of adjustable load. iz The calculation is shown in the following formula:

[0120]

[0121] In the formula p idz Let I be the average active power of the i-th adjustable load during the z-th dispatch period on the d-th day of the electricity consumption cycle; I(·) is a judgment function, which is 1 if the condition in the parentheses is true, and 0 otherwise.

[0122] The model building module 24 is used to take the types of adjustable loads, average power, and external information as observations, and to set electricity consumption rewards by combining the operating probability of adjustable loads in each scheduling period, thereby building and training an A2C-based optimal electricity consumption model for users.

[0123] Furthermore, the external information includes outdoor and indoor temperature information, time-of-use electricity pricing, historical start-stop data for each adjustable load, and user vehicle usage information; the electricity consumption rewards include electricity cost rewards, indoor temperature rewards, user comfort rewards, equipment operation restriction rewards, and electricity demand rewards. The equipment operation restriction rewards include operation restriction rewards for uninterrupted loads. The electricity demand rewards are given after the optimal electricity consumption model provides start-stop suggestions for adjustable loads. If the multi-task load monitoring model detects that the user adopts the corresponding suggestion, then each adjustable load receives a corresponding positive reward; otherwise, it receives a corresponding negative reward.

[0124] The electricity cost reward pri The calculation is shown in the following formula:

[0125] reward pri =-pri next ·p main

[0126] In the formula pri next The time-of-use electricity price at the next scheduling time, p main The active power input to the meter bus in the next scheduling time;

[0127] The indoor temperature reward tem The calculation is shown in the following formula:

[0128]

[0129] In the formula tem next The indoor temperature in Celsius for the next scheduling period;

[0130] The user comfort reward com The calculation is shown in the following formula:

[0131]

[0132] In the formula, act i For the action of the i-th adjustable load in the next scheduling period obtained from the user's optimal power consumption method, act i A value of 1 indicates that the i-th adjustable load is enabled, otherwise it is disabled; z refers to the corresponding index of the next scheduling period; I(·) is a judgment function, which is 1 if the condition in the parentheses is true, otherwise it is 0.

[0133] The optimal electricity consumption model for users includes an actor network and a critic network;

[0134] The optimization objective of the actor network is shown in the following formula:

[0135]

[0136] In the formula, b is the empirical number of parameters involved in gradient calculation in a single parameter update; For the qth experience, the observation during the z-th scheduling period Make an action The reward p represents the observation of the qth experience during the (z+1)th scheduling period; π (·) represents the parameterized model of the actor network, where π represents the corresponding model parameters; V φ (·) represents the parameterized model of the critic network, and φ represents the corresponding model parameters;

[0137] The optimization objective of the critic network is shown in the following formula:

[0138]

[0139] In the formula V φ' (·) represents the critic network before this parameter update, φ' represents the corresponding frozen model parameters, which are treated as constants during the optimization process; γ is the discount factor.

[0140] The model optimization module 25 is used to put the trained user-optimal electricity consumption model into application. When the multi-task load monitoring model detects a change in the average power or type of the user's adjustable load, it re-triggers the user-optimal electricity consumption model training process, updates the operating probability of the adjustable load in each time period and the type vector of the user's adjustable load, and modifies the electricity consumption reward.

[0141] The multi-task load monitoring and excitation power optimization device provided in this embodiment of the invention acquires a coarse-labeled meter dataset of adjustable loads, and preprocesses the coarse-labeled meter dataset to obtain a preprocessed meter dataset; constructs a multi-task load monitoring model, divides the preprocessed meter dataset into a training set and a validation set, and trains the multi-task load monitoring model on the training set until the loss function value of the multi-task load monitoring model converges on the validation set; deploys the trained multi-task load monitoring model to the user terminal, and analyzes the meter data of the power consumption cycle to obtain the types of adjustable loads of the user. This invention utilizes the types, average power, and operational probabilities of adjustable loads in each scheduling period. It takes the types, average power, and external information of user adjustable loads as observations, integrates the operational probabilities of adjustable loads in each scheduling period to set electricity consumption rewards, and constructs and trains an A2C-based user-optimal electricity consumption model. The trained user-optimal electricity consumption model is then put into application. When the multi-task load monitoring model detects a change in the average power or type of user adjustable load, the user-optimal electricity consumption model training process is retried, updating the operational probabilities of adjustable loads in each time period and the type vector of user adjustable loads, and modifying the electricity consumption rewards. Compared to existing technologies, this invention does not rely on expensive smart home platforms or depend on users providing appliance information and usage needs. It can be applied to incentive-based users who only have access to meter data, and can automatically detect changes and adjust strategies, meeting practical application requirements.

[0142] This invention also provides a computer-readable storage medium, which includes a stored computer program; wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the power optimization method of multi-task load monitoring and incentive as described above.

[0143] This invention also provides an electrical device, see [link to relevant documentation]. Figure 5 The diagram shown is a structural block diagram of a preferred embodiment of a power device provided by the present invention. The power device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power consumption optimization method of multi-task load monitoring and incentive as described above.

[0144] Preferably, the computer program can be divided into one or more modules / units (such as computer program 1, computer program 2, ...), and the one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the power equipment.

[0145] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor. The processor is the control center of the power equipment, connecting various parts of the power equipment through various interfaces and lines.

[0146] The memory mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard drive, a SmartMedia Card (SMC), a Secure Digital (SD) card, and a Flash Card, or other volatile solid-state storage devices.

[0147] It should be noted that the aforementioned power equipment may include, but is not limited to, processors and memory, as will be understood by those skilled in the art. Figure 5 The structural block diagram is merely an example of electrical equipment and does not constitute a limitation on the electrical equipment. It may include more or fewer components than shown, or combine certain components, or use different components.

[0148] In summary, the power consumption optimization method, apparatus, storage medium, and power equipment for multi-task load monitoring and incentive provided by this invention acquire a coarse-labeled meter dataset of adjustable loads, and preprocess the coarse-labeled meter dataset to obtain a preprocessed meter dataset; construct a multi-task load monitoring model, divide the preprocessed meter dataset into a training set and a validation set, and train the multi-task load monitoring model on the training set until the loss function value of the multi-task load monitoring model converges on the validation set; deploy the trained multi-task load monitoring model to the user terminal, and analyze the meter data of the electricity consumption cycle to obtain user... This invention identifies the types, average power, and operational probabilities of adjustable loads during each scheduling period. Using these factors, along with external information, as observations, and considering the operational probabilities of adjustable loads in each scheduling period, an A2C-based optimal user electricity consumption model is constructed and trained. The trained model is then applied. When the multi-task load monitoring model detects changes in the average power or type of adjustable loads, the optimal user electricity consumption model training process is re-triggered, updating the operational probabilities of adjustable loads in each time period and the type vector of adjustable loads, and modifying the electricity consumption rewards. Compared to existing technologies, this invention does not rely on expensive smart home platforms or depend on users providing appliance information and usage needs. It can be applied to incentive-driven users who only have access to meter data and can automatically detect changes and adjust strategies, meeting practical application requirements.

[0149] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A power consumption optimization method for multi-task load monitoring and excitation, characterized in that, The method includes the following steps: Obtain the coarse-labeled meter dataset of adjustable load, and preprocess the coarse-labeled meter dataset to obtain the preprocessed meter dataset; A multi-task load monitoring model is constructed. The preprocessed electricity meter dataset is divided into a training set and a validation set. The multi-task load monitoring model is trained on the training set until the loss function value of the multi-task load monitoring model converges on the validation set. The trained multi-task load monitoring model is deployed to the user end, and the meter data of the electricity consumption cycle is analyzed to obtain the types of adjustable loads, average power, and the probability of adjustable load operation in each scheduling period. By taking the types of adjustable loads, average power, and external information as observations, and combining the operating probability of adjustable loads in each scheduling period to set electricity consumption rewards, an A2C-based optimal electricity consumption model for users is constructed and trained. The trained optimal power consumption model is then deployed. When the multi-task load monitoring model detects a change in the average power or type of the user's adjustable load, the optimal power consumption model training process is retried. This re-updates the operating probabilities of the adjustable loads at different time periods and the type vectors of the user's adjustable loads, and modifies the power consumption rewards. These rewards include user comfort rewards. The calculation is shown in the following formula: In the formula The next scheduling period obtained from the user's optimal electricity consumption model. An adjustable load action, A value of 1 indicates the start of the first... If the load is adjustable, it will be off; otherwise, it will be off. For the first The first type of adjustable load The probability of adjustable load operation during each scheduling period; This refers to the index corresponding to the next scheduling period; This is a conditional function; if the condition within the parentheses is true, the value is 1; otherwise, it is 0.

2. The power consumption optimization method for multi-task load monitoring and excitation according to claim 1, characterized in that, The adjustable target load includes air conditioners, electric vehicles, and dishwashers; the coarse-labeled meter dataset is a dataset of active power from smart meters with timestamps of load device activation and deactivation. The coarsely labeled electricity meter dataset is preprocessed, specifically including: Generate a sequence of 0 and 1 states for each sampling point corresponding to each adjustable load based on the input and output timestamps; Calculate the mean active power in the coarsely labeled electricity meter dataset. with standard deviation The mean is subtracted from each sampling point and then divided by the variance to complete the data normalization operation. With width as The sliding step size is The sliding window divides the coarsely labeled electricity meter dataset and the switch state 0, 1 sequence into multiple segments. The average active power of each adjustable load is calculated within each segment based on the 0, 1 sequence of switching states. .

3. The power consumption optimization method for multi-task load monitoring and excitation according to claim 2, characterized in that, Constructing a multi-task load monitoring model specifically includes: The input fragmented meter data is processed by a one-dimensional convolutional layer to initially extract the intrinsic relationship between adjacent sampling points; The long-term operating mode of the load equipment is extracted by the bidirectional gated cyclic network, and then the average active power of various adjustable loads in the segment is mapped by the fully connected layers corresponding to various adjustable loads. Multiply the average active power of various adjustable loads by the standard deviation of active power in the coarse-labeled meter data set. In addition, the average active power in the coarsely labeled electricity meter dataset After completing the inverse normalization, a multi-task load monitoring model is obtained.

4. The power consumption optimization method for multi-task load monitoring and excitation according to claim 3, characterized in that, The loss function of the multi-task load monitoring model for: In the formula The number of types of adjustable loads. For the first Number of sample segments for adjustable loads For the first The first type of adjustable load The average active power value of the sample segments. For the first The first type of adjustable load The active power sequence after normalization of the sample segments. This refers to the constructed multi-task load monitoring model.

5. The power consumption optimization method for multi-task load monitoring and excitation according to claim 4, characterized in that, The trained multi-task load monitoring model is deployed to the user end, and meter data during the electricity consumption cycle is analyzed to obtain the types of adjustable loads, average power, and the probability of adjustable load operation during each scheduling period. Specifically, this includes: The preprocessed meter dataset is input into the multi-task load monitoring model to obtain the average active power of various adjustable loads in each segment; The average active power of each segment is calculated to obtain the total average active power of various adjustable loads in the user's household. ; Determine the total average active power If it is greater than the threshold If the user's adjustable load type vector is true, then the corresponding position is set to 1; otherwise, it is set to 0. Where, the first... The first type of adjustable load Adjustable load operation probability for each scheduling period The calculation is shown in the following formula: In the formula For the first Adjustable load in the first cycle of electricity consumption Tianzhongdi The average active power within each scheduling period, where t is the number of days in the electricity consumption cycle; This is a conditional function; if the condition within the parentheses is true, the value is 1; otherwise, it is 0.

6. The power consumption optimization method for multi-task load monitoring and excitation according to claim 5, characterized in that, The external information includes outdoor and indoor temperature information, time-of-use electricity pricing, historical start-stop data for each adjustable load, and user vehicle usage information; the electricity consumption rewards also include electricity cost rewards, indoor temperature rewards, equipment operation restriction rewards, and electricity demand rewards. The equipment operation restriction rewards include rewards for uninterrupted load operation restrictions. The electricity demand rewards are given after the optimal electricity consumption model provides start-stop suggestions for adjustable loads. If the multi-task load monitoring model detects that the user has adopted the corresponding suggestion, each adjustable load receives a corresponding positive reward; otherwise, it receives a corresponding negative reward. Electricity cost incentive The calculation is shown in the following formula: In the formula The time-of-use electricity price at the next scheduling time. The active power input to the meter bus in the next scheduling time; The indoor temperature reward The calculation is shown in the following formula: In the formula This is the indoor temperature value in Celsius for the next scheduling period.

7. The power consumption optimization method for multi-task load monitoring and excitation according to claim 1, characterized in that, The optimal electricity consumption model for users includes an actor network and a critic network; The optimization objective of the actor network is shown in the following formula: In the formula The empirical number of gradients involved in a single parameter update; For the first The first experience Observations during each scheduling period Make an action The reward Indicates the first The first experience Observations during each scheduling period; A parameterized model for actor networks. Indicates the corresponding model parameters; A parameterized model for the critic network. Indicates the corresponding model parameters; The optimization objective of the critic network is shown in the following formula: In the formula This refers to the commentator network before this parameter update. This represents the corresponding frozen model parameters, which are treated as constants during the optimization process; This is the discount factor.

8. A power consumption optimization device for multi-task load monitoring and excitation, characterized in that, The device includes: The data preprocessing module is used to obtain the coarse-labeled meter dataset of adjustable load and preprocess the coarse-labeled meter dataset to obtain the preprocessed meter dataset. The model training module is used to construct a multi-task load monitoring model, divide the preprocessed meter dataset into a training set and a validation set, and train the multi-task load monitoring model on the training set until the loss function value of the multi-task load monitoring model on the validation set converges. The data analysis module is used to deploy the trained multi-task load monitoring model to the user end and analyze the meter data of the electricity consumption cycle to obtain the types of adjustable loads, average power, and the probability of adjustable load operation in each scheduling period. The model building module is used to take the types of adjustable loads, average power, and external information as observations, and to set electricity consumption rewards by combining the operating probability of adjustable loads in each scheduling period, thereby building and training an A2C-based optimal electricity consumption model for users. The model optimization module is used to apply the trained user-optimal electricity consumption model. When the multi-task load monitoring model detects a change in the average power or type of the user's adjustable load, it re-triggers the user-optimal electricity consumption model training process, updates the operating probabilities of the adjustable load at each time period and the type vector of the user's adjustable load, and modifies the electricity consumption reward, which includes a user comfort reward. The calculation is shown in the following formula: In the formula The next scheduling period obtained from the user's optimal electricity consumption model. An adjustable load action, A value of 1 indicates the start of the first... If the load is adjustable, it will be off; otherwise, it will be off. For the first The first type of adjustable load The probability of adjustable load operation during each scheduling period; This refers to the index corresponding to the next scheduling period; This is a conditional function; if the condition within the parentheses is true, the value is 1; otherwise, it is 0.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program; wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the power consumption optimization method of multi-task load monitoring and stimulation as described in any one of claims 1 to 7.

10. An electrical device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the power optimization method for multi-task load monitoring and stimulating as described in any one of claims 1 to 7.

Images