Lstm-based load aggregator modeling method under user response mechanism

Abstract

The application belongs to the technical field of power system dispatching optimization, and more particularly relates to a robust optimization-based method for modeling a load aggregation commercial power model under a user response mechanism in an elastic environment. First, response models of various users are analyzed and established, historical weather, illumination and user load baseline data are taken as inputs of the user response models, so as to obtain data samples of the users under different UDR instructions. Then, the users are classified according to their incentive types, and the response data of the users of the same type are aggregated to obtain a training sample set. Finally, a corresponding LSTM model is constructed, and the LSTM model is trained according to the training sample, so as to reduce the model representation error.

Description

Technical Field

[0001] This invention belongs to the field of demand response potential prediction, and more specifically, relates to a load aggregator modeling method based on LSTM under the user response mechanism. Background Technology

[0002] In recent years, with the continuous growth of electricity demand and changes in load structure, the peak-to-valley difference in electricity consumption has been widening year by year, and seasonal power shortages occur frequently. Relying solely on generation-side regulation is insufficient to ensure a balance between power supply and demand. Therefore, it is necessary to mobilize demand-side resources to participate in power regulation and alleviate power supply pressure. Currently, the grid side generally uses demand response (DR) to regulate demand, but this is generally only applicable to large industrial users and load aggregators (LAs), leaving the demand response potential of small and medium-sized users at the grassroots level untapped. As power agents, LAs can guide grassroots users to adjust their electricity consumption behavior by implementing user demand response, thereby responding to grid demand. Once LAs accurately grasp the response capabilities of grassroots users, they can formulate reasonable user demand response (UDR) policies tailored to these users, improving their own economic efficiency and user participation, and enhancing the reliability of the response. Therefore, it is crucial for LAs to establish their electricity consumption models, that is, to understand the overall response behavior of all users considering their own primary objectives and under different incentive conditions of the LA.

[0003] Due to the inherent uncertainty in user energy consumption, and the varying incentive and response mechanisms among different users, it is difficult to obtain comprehensive data on the overall electricity consumption behavior of the integrated energy system (LA). Currently, scholars have conducted research on LA electricity consumption modeling. Some scholars have established a temperature-controlled load aggregation model that considers user response uncertainty based on the physical model and statistical principles of a single temperature-controlled load. Others have introduced demand price elasticity, analyzed the specific modeling process of self-elasticity and cross-elasticity, and proposed a load response model for large electricity users under time-of-use pricing. Some scholars have used the interval method to characterize the uncertainty of the integrated energy system (DR) of the community, but because it only uses the boundary of user response capabilities to represent the uncertainty of their response, the characterization accuracy is low. The establishment of the above models requires detailed user electricity consumption data to build a relatively accurate energy consumption model. However, in reality, there are numerous users at the grassroots level with varying response capabilities, making it difficult to obtain a unified model. Furthermore, this data involves user privacy and is difficult to obtain.

[0004] Furthermore, some scholars have used data-driven approaches to model LA (Local Area) electricity consumption, building models based on historical electricity usage data of internal users. Others have established incentive-based user response models using time-series decomposition methods, based on user participation rates and effective response rates. Semi-supervised learning is employed to identify the demand response potential of individual households under dynamic electricity pricing. Still others have constructed a user energy consumption model under electricity price incentives based on time-series networks, utilizing historical user data, electricity prices, temperature, and humidity data. However, in the aforementioned studies, LAs implement a single incentive method for all internal users; in reality, LAs often contain multiple user types, and using the same incentive method for different users cannot fully stimulate their demand response potential. The large number of user types leads to excessively large data dimensions, posing new challenges to LA modeling. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a load aggregator modeling method based on LSTM under the user response mechanism, which can effectively predict the response capability of LA.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A load aggregator modeling method based on LSTM under a user response mechanism, characterized by the following steps:

[0008] Step 1: Analyze the user response characteristics within LA, establish user response models, and generate user response data;

[0009] Step 2: Aggregate user data based on user incentive type to generate a data sample set;

[0010] Step 3: Construct the corresponding LSTM network model and train the model based on the training samples.

[0011] This technical solution is further optimized so that the LA (Local Area) consists of an energy management center and various types of users. When the LA receives a DR (Demand Response) instruction from the power grid, the LA's energy management center will simulate user response behavior based on a constructed electricity consumption model and formulate a reasonable UDR (User Demand Response) policy considering its own economics to incentivize underlying users to adjust their electricity demand, thereby responding to the power grid's demand. The LA includes four types of users: residential users, commercial building users, charging stations, and hospitals. The electricity demand of commercial buildings, charging stations, and hospitals is controlled by their respective operators as a whole, and their response capabilities are relatively stable. Generally, incentive-based user demand response is used for regulation. Communities consist of numerous household units, and the electricity consumption habits of each household vary greatly, making direct regulation difficult. Generally, the difference between peak and off-peak electricity prices and peak hours is adjusted to guide users to consume electricity in an orderly manner.

[0012] Further optimization of this technical solution is as follows: The user response models in step 1 are shown below:

[0013] Step 1.1: Construct a resident user response model:

[0014] The residential user unit includes rigid equipment, household photovoltaic equipment, and adjustable equipment. Adjustable equipment can be further categorized into reduceable equipment, transferable equipment, and repositionable equipment based on its adjustment method. Its response model can be expressed as:

[0015]

[0016] in, c ele (k) represents the dissatisfaction cost of the load that can be reduced, the electricity demand, and the residential electricity price for the k-th time period.

[0017] Step 1.2: Construct a user response model for commercial buildings:

[0018] The interior of the commercial building-type user mainly includes air conditioning and some rigid equipment, and the load reduction during peak hours can be achieved by adjusting the air conditioning power. Its response model can be expressed as:

[0019]

[0020] in, ε air (k) These represent the original power consumption, actual power consumption, subsidy price, and dissatisfaction cost caused by load reduction for the air conditioner in the k-th time period.

[0021] Step 1.3: Construct the hospital's response model:

[0022] Hospitals' electricity demand comes from medical equipment, elevators, lighting, air conditioning, and other facilities, requiring high reliability of power supply. Hospital administrators can reduce load by adjusting the power consumption of equipment such as air conditioners and water heaters. However, compared to commercial users, hospitals have more rigid equipment, resulting in a lower percentage of electricity consumption that can be reduced and a weaker response capability. Its response model can be expressed as:

[0023]

[0024] in, ε cut (k) These represent the original power consumption of the equipment that can be reduced in the k-th time period, the actual power consumption, the subsidy price, and the dissatisfaction cost caused by the load reduction.

[0025] Step 1.4: Construct the response model of the charging station:

[0026] The primary load demand of charging stations comes from customer charging needs. After receiving the invitation information, charging station operators adjust their service electricity prices for peak, off-peak, and flat periods to guide customers to adjust their charging times, thus achieving load shifting. Its response model can be expressed as:

[0027]

[0028] in, c serve (k), ε tr (k) These are the power purchased by the charging station, the original power purchased, the retail electricity price, the load aggregator subsidy price, and the output of the photovoltaic equipment.

[0029] Further optimization of this technical solution, the data classification and aggregation process in step 2 is as follows:

[0030] Step 2.1: Price-based user sample aggregation:

[0031] For price-sensitive users such as residential users, a set of environmental data e = (H, G) is selected from the collected historical environmental dataset E, including the temperature curve H = (h(0), h(1), ..., h(K-1)). T And the light intensity curve G=(g pv (0),g pv (1),…,g pv (K-1) T The selected environmental data e and the user's historical load baseline under environmental data e will be used. and the time-of-use electricity pricing C released by LA ele =(c ele (0),c ele (1),…c ele (K-1) T As a factor influencing the user's energy consumption, its characteristic matrix was obtained. By inputting its feature matrix into the user's optimization model, the user's response curve at a certain response level can be obtained. Use its response curve as a sample label A group As a sample pair, different environmental data are extracted, and the time-of-use electricity price is adjusted to obtain the user's response curve under different responsiveness levels, generating I. m The group of samples constitutes the user's sample set

[0032] Step 2.2: Aggregation of Incentive-Based User Samples:

[0033] For incentive-driven users such as commercial buildings, charging stations, and hospitals, a set of environmental data is extracted from dataset E, and the environmental data and the user load baseline in that environment are combined. And the compensation electricity price ε issued by LA to users n =(ε n (0),…,ε n (K-1) T As the energy consumption characteristics of this user, its feature matrix is ​​obtained. By inputting its feature matrix into the user's optimization model, the user's response curve under the current stimulus can be obtained. As a sample label A group As a sample pair, different environmental data are extracted, and the compensation electricity price is adjusted to obtain the user's response curve under different responsiveness, generating I. n The group of samples together constitutes the user's training sample set.

[0034] Step 2.3: Data Sample Aggregation:

[0035] The LA electricity consumption model is aggregated from the response behaviors of each user. A set of sample pairs under the same environmental data 'e' is selected from the sample sets of each user, and the load baselines of all price-based users are superimposed to form the overall load baseline for all price-based users. The load baselines of all incentivized users in the sample are summed to form the overall load baseline for all incentivized users. The instruction issued by LA to the user is U = {C ele ,ε n},Will This serves as the power consumption characteristic matrix for the LA. The response curves of each user are superimposed to form the power consumption response curve of the LA after implementing UDR. As a sample label This yields a set of samples after LA implements UDR in the current environment. Adjusting the UDR instruction to generate I under different environments s Grouping samples to obtain a sample set

[0036] Further optimization of this technical solution, the LSTM network structure and training process described in step 3 are as follows:

[0037] Step 3.1: LSTM network structure:

[0038] LSTM consists of a forget gate, an input gate, and an output gate, based on the output h from the previous time step. k-1 C k-1 and the state characteristics of the input at the current moment Get the output h at the next time step k C kThe forget gate determines how much information from the previous time step is retained. The closer the forget gate output is to 0, the more information is discarded; conversely, the closer it is to 1, the more information is retained. Its update formula is shown below:

[0039]

[0040] The input gate determines whether the input information at the current moment is retained in the current state, and its update formula is as follows:

[0041]

[0042] The output gate determines how much of the current state is output to the next time step, and its update formula is as follows:

[0043]

[0044] in, Let W be the feature data for the k-th time period, σ be the activation function, and W be the value of W. f W i W c W o Let b be the weight matrix. f ,b i ,b c ,b o As a bias, the model output Will The response curve is obtained after inverse normalization.

[0045] Step 3.1: LSTM Network Training:

[0046] Because the scales of various types of information are inconsistent, it is necessary to transform the feature matrix x s and tag data Normalization is performed:

[0047]

[0048] in, For the feature data in the k-th row and j-th column, The minimum and maximum values ​​in column j are the values ​​for the label data, which can be obtained similarly.

[0049] After normalizing the dataset, the hold-out method is used to divide it into three mutually exclusive sets in a 6:2:2 ratio, serving as the training set, test set, and validation set, respectively. The training and validation sets are used for model training, while the test set is used to evaluate the model's performance.

[0050] The network weights and biases are trained using a stochastic optimization method with adaptive momentum. During training, the root mean square error (RMSE) function is used as the loss function for LSTM training to measure the similarity between the predicted and labeled values, as shown below:

[0051]

[0052] Compared with the prior art, the beneficial effects of the present invention are mainly reflected in:

[0053] This invention fully evaluates the demand response capability of energy-consuming applications (LAs). Addressing the issue of excessive data dimensionality due to the large number of LA users, it categorizes users according to their incentive methods and aggregates their feature data. After aggregation, LSTM is used to model the electricity consumption behavior of LAs. This invention reduces the parameter dependence of the LA energy consumption model and lowers its complexity. Attached Figure Description

[0054] Figure 1 Implement an architecture diagram to respond to user needs;

[0055] Figure 2 Implementation architecture diagram for responding to incentivized user needs;

[0056] Figure 3 Flowchart for modeling the electricity consumption model;

[0057] Figure 4 This is a diagram of the LSTM network structure. Detailed Implementation

[0058] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0059] This invention discloses a load aggregator (UDR) modeling method based on LSTM under a user response mechanism. The load aggregator (LA) consists of an energy management center and various users. When the LA receives a DR instruction from the grid, the LA's energy management center simulates user response behavior based on a constructed electricity consumption model and formulates a reasonable UDR policy considering its own economic viability. This incentivizes underlying users to adjust their electricity demand, thereby responding to grid demand. Specifically, as follows... Figure 1 As shown. The LA includes four types of users: residential users, commercial building users, charging stations, and hospitals. The electricity demand of commercial buildings, charging stations, and hospitals is controlled by their respective operators as a whole, and their response capabilities are relatively stable. Generally, incentive-based user demand response is used for regulation. Communities consist of many household units, and the electricity consumption habits of each household vary greatly, making direct regulation difficult. Generally, the difference between peak and off-peak electricity prices and peak hours is adjusted to guide users to use electricity in an orderly manner.

[0060] The price-based user demand response (UTR) described in this invention primarily guides user electricity consumption behavior through economic levers to achieve load shifting and load reduction. It mainly includes time-of-use pricing (TOU), real-time pricing, and peak-hour pricing. Currently, real-time pricing is not yet mature in my country, while TOU, as a simple and effective UDR method, has been promoted and is gradually becoming more widespread both domestically and internationally.

[0061] The incentive-based UDR described in this invention refers to a LA (Local Authority) providing compensation or preferential electricity prices to users who reduce their electricity demand during specific periods. It is mainly divided into planned and market-based UDRs. This paper primarily studies the response behavior of users under a planned UDR, and the specific implementation process is as follows: Figure 2 As shown. LA predicts user response capabilities based on load baseline and weather forecast information, and, considering its own economics, sends invitation information {ε,Φ} to underlying contracted users. LA The subsidized electricity price ε(k) is shown below:

[0062]

[0063] In the formula: ε is the unit subsidy price given by LA, Φ LA This is the period during which the LA (Local Authority) requests load adjustments from users. Users will consider their own economic and comfort considerations based on the invitation information to decide the amount to participate in the bidding and provide feedback to the LA. The LA will then notify each user of the successful or unsuccessful bid and determine the amount each user needs to respond to. The following day, the LA will notify participating users at the agreed time whether to implement the UDR (Usage-Based Delivery) as agreed. After the response is completed, users will receive compensation according to the contract. Different subsidy rates are applied to different users.

[0064] This embodiment proposes a load aggregator modeling method based on LSTM under the user response mechanism, which includes the following steps:

[0065] Step 1: Analyze the user response characteristics within LA, establish user response models, and generate user response data.

[0066] The user response models for step 1 are shown below:

[0067] Step 1.1: Construct a resident user response model:

[0068] Residential users' equipment includes rigid equipment, household photovoltaic equipment, and adjustable equipment. Adjustable equipment can be divided into reduceable equipment, transferable equipment, and lateral equipment according to the different adjustment methods.

[0069] Reduceable power equipment refers to equipment whose power can be adjusted within a certain range while meeting the user's electricity demand. As the equipment power deviates from the optimal or rated operating power, user dissatisfaction will occur, and this dissatisfaction will have consequences. It can be represented as:

[0070]

[0071] in, These represent the power consumption that can be reduced and the optimal operating power of the equipment during the k-th time period, respectively. This represents the dissatisfaction coefficient. Equipment operation can be reduced while meeting upper and lower power limits:

[0072]

[0073] in, These are the lower and upper power limits of the device that can be reduced in the k-th time period, respectively.

[0074] Relocatable equipment refers to equipment that, once started, cannot be stopped or whose stopping cost is too high; generally, only its start-up time can be adjusted. The dissatisfaction and costs caused by relocating equipment during its power consumption period are considered.

[0075]

[0076] Start-up time of movable equipment Constraints must be met:

[0077]

[0078] in, These are the earliest start time, latest start time, and initial start time of the movable equipment, respectively. The discomfort coefficient of a load that can be transferred.

[0079] Portable equipment refers to equipment whose power consumption periods and power consumption can be adjusted by the user, while the total power load remains unchanged throughout the day.

[0080]

[0081] in, Let be the power consumption of the portable device in the k-th time period. Baseline power for portable devices.

[0082] Upper and lower limits of operating power for transferable equipment:

[0083]

[0084] in, These represent the minimum and maximum operating power during the k-th time period after the transfer, respectively.

[0085] Residential electricity consumption should meet the purchase and demand balance constraint:

[0086]

[0087] Among them, P rigid (k) This represents the average power consumption, user-purchased power, and photovoltaic (PV) equipment output for rigid equipment, portable equipment, movable equipment, and equipment that can be reduced in size during the k-th time period. PV equipment output is related to the solar irradiance G. pv (k) related.

[0088] The response model for residential users can be represented as:

[0089]

[0090] in, c ele (k) represents the dissatisfaction cost of load reduction in the k-th time period and the residential electricity price, respectively.

[0091] Step 1.2: Construct a user response model for commercial buildings:

[0092] The interior of the commercial building-type user mainly includes air conditioning and some rigid equipment, and the load reduction during peak hours can be achieved by adjusting the air conditioning power. Its response model can be expressed as:

[0093]

[0094] in, ε air (k) These represent the original power consumption, actual power consumption, subsidy price, and dissatisfaction cost caused by load reduction for the air conditioning in the k-th time period. The dissatisfaction cost is calculated as follows:

[0095]

[0096] in, To reduce the optimal operating power of the equipment, the dissatisfaction cost coefficient is...

[0097] The equipment operation should meet the upper and lower power limits:

[0098]

[0099] in, These represent the minimum and maximum operating power that can be reduced for the equipment during time period k, respectively.

[0100] Step 1.3: Construct the hospital's response model:

[0101] Hospitals' electricity demand comes from medical equipment, elevators, lighting, air conditioning, and other facilities, requiring high reliability of power supply. Hospital administrators can reduce load by adjusting the power consumption of equipment such as air conditioners and water heaters. However, compared to commercial users, hospitals have more rigid equipment, resulting in a lower percentage of electricity consumption that can be reduced and a weaker response capability. Its response model can be expressed as:

[0102]

[0103] in, ε cut (k) These represent the original power consumption, actual power consumption, subsidy price, and dissatisfaction cost caused by load reduction for the equipment that can be reduced in the k-th time period. The dissatisfaction cost is calculated as follows:

[0104]

[0105] in, To reduce the optimal operating power of the equipment, the dissatisfaction cost coefficient is...

[0106] The equipment operation should meet the upper and lower power limits:

[0107]

[0108] In the formula: These represent the minimum and maximum operating power that can be reduced for the equipment during time period k, respectively.

[0109] Step 1.4: Construct the response model of the charging station:

[0110] The primary load demand for charging stations comes from customer charging needs. After receiving invitations, charging station operators adjust retail electricity prices for peak, off-peak, and flat periods to guide customers to adjust their charging times, thus achieving load shifting. The retail electricity price is determined by the base price c. base (k) Superimposed service electricity price c serve (k) constitutes, c serve (k) Generally, the TOU method is adopted, and its peak and valley periods are divided in accordance with c. base The peak and valley times are divided in the same way for (k). Charging station operators adjust c... serve (k) Electricity prices at different times guide customer load shifting; based on consumer psychology, the peak-valley service electricity price difference Δc can be obtained. hv With peak-valley transition rate r hv The relationship between them is as follows:

[0111]

[0112] in, The peak-valley transition rate coefficient, d hv This is a triangular fuzzy number, representing the uncertainty of customer response and... Related to the peak-valley electricity price difference:

[0113]

[0114] in, This is the peak-to-valley transfer rate coefficient. Similarly, based on the peak-to-average transfer rate coefficient and the average-to-valley transfer rate coefficient, the peak-to-average transfer rate r can be obtained. hf Pinggu transfer rate r fv The electricity purchased by the customer in the k-th time period after the response can be expressed as:

[0115]

[0116] in, Φ represents the original power consumption in the k-th time period. h ,Φ f ,Φ v P represents the set of three time periods: peak, trough, and flat. h-a P f-a These represent the average power consumption during peak and off-peak hours, respectively, which can be expressed as:

[0117]

[0118] Among them, l h ,l f These represent the number of unit times that the peak and off-peak periods last, respectively.

[0119] The response model of a charging station can be expressed as:

[0120]

[0121] in, c serve (k), ε tr (k) These are the power purchased by the charging station, the original power purchased, the retail electricity price, the load aggregator subsidy price, and the output of the photovoltaic equipment.

[0122] Step 2: Aggregate user data based on user incentive type to generate a data sample set.

[0123] Step 2: Data classification and aggregation process as follows Figure 3 As shown:

[0124] Step 2.1: Price-based user sample aggregation:

[0125] For price-sensitive users such as residential users, a set of environmental data e = (H, G) is selected from the collected historical environmental dataset E, including the temperature curve H = (h(0), h(1), ..., h(K-1)). T And the light intensity curve G=(g pv (0),g pv (1),…,g pv (K-1) T The selected environmental data e and the user's historical load baseline under environmental data e will be used. and the time-of-use electricity pricing C released by LA ele =(c ele (0),c ele (1),…c ele (K-1) T As a factor influencing the user's energy consumption, its characteristic matrix was obtained. By inputting its feature matrix into the user's optimization model, the user's response curve at a certain response level can be obtained. Use its response curve as a sample label A group As a sample pair, different environmental data are extracted, and the time-of-use electricity price is adjusted to obtain the user's response curve under different responsiveness levels, generating I. m The group of samples constitutes the user's sample set

[0126] Step 2.2: Aggregation of Incentive-Based User Samples:

[0127] For incentive-driven users such as commercial buildings, charging stations, and hospitals, a set of environmental data is extracted from dataset E, and the environmental data and the user load baseline in that environment are combined. And the compensation electricity price ε issued by LA to users n =(ε n (0),…,ε n (K-1) T As the energy consumption characteristics of this user, its feature matrix is ​​obtained. By inputting its feature matrix into the user's optimization model, the user's response curve under the current stimulus can be obtained. As a sample label A group As a sample pair, different environmental data are extracted, and the compensation electricity price is adjusted to obtain the user's response curve under different responsiveness, generating I. n The group of samples together constitutes the user's training sample set.

[0128] Step 2.3: Data Sample Aggregation:

[0129] The LA electricity consumption model is aggregated from the response behaviors of each user. A set of sample pairs under the same environmental data 'e' is selected from the sample sets of each user, and the load baselines of all price-based users are superimposed to form the overall load baseline for all price-based users. The load baselines of all incentivized users in the sample are summed to form the overall load baseline for all incentivized users. The instruction issued by LA to the user is U = {C ele ,ε n},Will This serves as the power consumption characteristic matrix for the LA. The response curves of each user are superimposed to form the power consumption response curve of the LA after implementing UDR. As a sample label This yields a set of samples after LA implements UDR in the current environment. Adjusting the UDR instruction to generate I under different environments s Grouping samples to obtain a sample set

[0130] Step 3: Construct the corresponding LSTM network model and train the model based on the training samples.

[0131] The LSTM network structure and training process described in step 3 are as follows:

[0132] Step 3.1: The LSTM network structure is as follows Figure 4 As shown:

[0133] LSTM consists of a forget gate, an input gate, and an output gate, based on the output h from the previous time step. k-1 C k-1 and the state characteristics of the input at the current moment Get the output h at the next time step k C k The forget gate determines how much information from the previous time step is retained. The closer the forget gate output is to 0, the more information is discarded; conversely, the closer it is to 1, the more information is retained. Its update formula is shown below:

[0134]

[0135] The input gate determines whether the input information at the current moment is retained in the current state, and its update formula is as follows:

[0136]

[0137] The output gate determines how much of the current state is output to the next time step, and its update formula is as follows:

[0138]

[0139] in, Let W be the feature data for the k-th time period, σ be the activation function, and W be the value of W. f W i W c W o Let b be the weight matrix. f ,b i ,b c ,b o For bias, the model output Will The response curve is obtained after inverse normalization.

[0140] Step 3.1: LSTM Network Training:

[0141] Because the scales of various types of information are inconsistent, it is necessary to transform the feature matrix x s and tag data Normalization is performed:

[0142]

[0143] in, For the feature data in the k-th row and j-th column, The minimum and maximum values ​​in column j are the values ​​for the label data, which can be obtained similarly.

[0144] After normalizing the dataset, the hold-out method is used to divide it into three mutually exclusive sets in a 6:2:2 ratio, serving as the training set, test set, and validation set, respectively. The training and validation sets are used for model training, while the test set is used to evaluate the model's performance.

[0145] The network weights and biases are trained using a stochastic optimization method with adaptive momentum. During training, the root mean square error (RMSE) function is used as the loss function for LSTM training to measure the similarity between the predicted and labeled values, as shown below:

[0146]

[0147] Compared with the prior art, the beneficial effects of the present invention are mainly reflected in:

[0148] This invention fully evaluates the demand response capability of energy-consuming applications (LAs). Addressing the issue of excessive data dimensionality due to the large number of LA users, it categorizes users according to their incentive methods and aggregates their feature data. After aggregation, LSTM is used to model the electricity consumption behavior of LAs. This invention reduces the parameter dependence of the LA energy consumption model and lowers its complexity.

[0149] This invention fully assesses the demand responsiveness of energy management systems (LAs). Addressing the issue of excessive data dimensionality due to the large number of users within an LA, it categorizes users according to their incentive methods and aggregates their feature data. After aggregation, LSTM is used to model the electricity consumption behavior of the LA. This invention reduces the parameter dependence of the LA energy consumption model and lowers its complexity.

[0150] It should be noted that, in this document, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Unless otherwise specified, an element defined by the phrase "comprising..." or "including..." does not exclude the presence of additional elements in the process, method, article, or terminal device that includes said element. Furthermore, in this document, "greater than," "less than," "exceeding," etc., are understood to exclude the stated number; "above," "below," "within," etc., are understood to include the stated number.

[0151] Although the above embodiments have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the above descriptions are merely embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A modeling method for commercial power load aggregation based on LSTM under a user response mechanism, characterized in that, Includes the following steps: Step 1: Analyze the user response characteristics within the load aggregator, establish user response models, and generate user response data; Step 2: Aggregate user response data based on user incentive type to generate a data sample set; The data aggregation process in step 2 is as follows: Step 2.1: Price-based user sample aggregation For price-sensitive users, such as residential users, the collected historical environmental datasets are used... Select a set of environmental data This includes temperature profiles. Light intensity curve Selected environmental data User in environmental data Historical load baseline Time-of-use pricing published by load aggregators As a factor influencing the user's energy consumption, its characteristic matrix was obtained. The feature matrix is ​​then input into the user's optimization model to obtain the user's response curve at a certain response level. Use its response curve as the sample label A group As a sample pair, different environmental data are extracted, and the time-of-use electricity price is adjusted to obtain the user's response curve under different responsiveness levels, generating... The group of samples constitutes the user's sample set ; Step 2.2: Aggregation of Incentive-Based User Samples For incentive-based users such as commercial buildings, charging stations, and hospitals, from the dataset Extract a set of environmental data, including the environmental data and the user load baseline under that environment. Compensation rates for users issued by load aggregators As a characteristic of the user's energy consumption, we obtain Its characteristic matrix The user's response curve under the current stimulus is obtained by inputting its feature matrix into the user's optimization model. As a sample label A group As a sample pair, different environmental data are extracted, and the compensation electricity price is adjusted to obtain the user's response curve under different responsiveness levels, generating... The group of samples together constitutes the user's sample set. ; Step 2.3: Data Sample Aggregation The load aggregator's electricity consumption model is formed by aggregating the response behaviors of each user, and selects a set of identical environmental data from the sample sets of each user. The sample pairs below are used to superimpose the load baselines of all price-based users to form the overall load baseline for all price-based users. The load baselines of all incentivized users in the sample are superimposed to form the overall load baseline for all incentivized users. The instructions issued by the load aggregator to the user are ,Will The electricity consumption characteristic matrix of the load aggregator is used to superimpose the response curves of each user to form the electricity consumption response curve of the load aggregator after implementing User Demand Response (UDR). As a sample label This yields a sample of load aggregators implementing UDR under the current environment. Adjusting UDR instruction generation under different environments Group samples to obtain a data sample set ; Step 3: Construct the corresponding LSTM network model and train the model based on the training samples.

2. The method for modeling a load aggregation commercial power model based on LSTM under the user response mechanism as described in claim 1, characterized in that, The load aggregator consists of an energy management center and various types of users. When the load aggregator receives a demand response instruction from the power grid, its energy management center will simulate the user's response behavior based on a constructed electricity consumption model and formulate a reasonable user demand response policy considering its own economics. This policy incentivizes lower-level users to adjust their electricity consumption to respond to the power grid's demand. The load aggregator includes four types of users: residential users, commercial building users, charging stations, and hospitals. The electricity demand of commercial buildings, charging stations, and hospitals is controlled by their respective operators, resulting in relatively stable response capabilities. Incentive-based user demand response is used for regulation. The community consists of numerous family units, and the electricity consumption habits of each family are quite different, making it difficult to directly regulate. The community guides users to use electricity in an orderly manner by adjusting the difference between peak and off-peak electricity prices and the off-peak hours.

3. The method for modeling a commercial power load aggregation model based on LSTM under the user response mechanism as described in claim 1, characterized in that, The user response models in step 1 are shown below: Step 1.1: Construct a residential user response model The residential user unit includes rigid equipment, household photovoltaic equipment, and adjustable equipment. Adjustable equipment is further categorized into reduceable equipment, transferable equipment, and movable equipment based on its adjustment method. Its response model is expressed as follows: in, , , , The first Dissatisfaction costs of load reduction for each time period, dissatisfaction costs of load shifting, electricity purchase demand, and residential electricity prices; Step 1.2: Construct a user response model for commercial buildings The commercial building type of user includes air conditioning and some rigid equipment. Peak-hour load reduction is achieved by adjusting the air conditioning power. Its response model is expressed as follows: in, , , , , The first The original power consumption, actual power consumption, subsidy price, basic electricity price, and unsatisfactory costs caused by load reduction for each time period of the air conditioner; Step 1.3: Constructing the hospital's response model Hospitals have a large amount of rigid equipment, and the proportion of electricity consumption that can be reduced from equipment is relatively low, resulting in poor responsiveness. Their response model is expressed as follows: in, , , , The first The original power consumption, actual power consumption, subsidy price, and dissatisfaction costs caused by load reduction can be reduced in a given period of time. Step 1.4: Construct the response model of the charging station The main load demand of charging stations comes from customers' charging needs. After receiving the invitation information, the charging station operator adjusts the service electricity price for peak, off-peak, and flat periods to guide customers to adjust their charging times, thereby achieving load shifting. Its response model is represented as follows: in, , , , , These are the power purchased by the charging station, the original power purchased, the retail electricity price, the load aggregator subsidy price, and the output of the photovoltaic equipment.

4. The method for modeling a commercial power load aggregation model based on LSTM under the user response mechanism as described in claim 1, characterized in that, The LSTM network structure and training process in step 3 are shown below: Step 3.1: LSTM Network Structure LSTM consists of a forget gate, an input gate, and an output gate, which calculate the output based on the previous time step. And the input at the current moment Get the output at the next time step The forget gate determines how much information from the previous state is retained. The closer the forget gate output is to 0, the more information is discarded; conversely, the closer it is to 1, the more information is retained. Its update formula is as follows: The input gate determines whether the input information at the current moment is retained in the current state, and its update formula is as follows: The output gate determines how much of the current state is output to the next time step, and its update formula is as follows: in, For the first Feature data for each time period For activation function, This is the weight matrix. For bias, the model output ,Will The response curve is obtained after inverse normalization. Step 3.2: LSTM Network Training Because various types of information have different scales, it is necessary to transform the feature matrix. and tag data Normalization is performed: in, For the first Line number Column feature data, For the first The minimum and maximum values ​​of the column can be obtained similarly for the label data; After the dataset is normalized, it is divided into three mutually exclusive sets in a 6:2:2 ratio using the hold-out method, which serve as the training set, test set, and validation set, respectively. The training set and validation set are used for training the model, and the test set is used to evaluate the performance of the model. The network weights and biases are trained using an adaptive momentum stochastic optimization method. During training, the root mean square error function is used as the loss function for LSTM training to measure the similarity between the predicted and labeled values, as shown below: 。

Images