Response potential evaluation method and system based on temperature-sensitive characteristics and non-parametric transfer learning

Abstract

The invention discloses a response potential evaluation method and system based on temperature-sensitive characteristics and non-parametric transfer learning. Relates to the technical field of power system automation. The method comprises the following steps: firstly, establishing a commercial user typical temperature-sensitive load feature model, and screening out source domain users matched with a potential increment user energy consumption mode based on a temperature-sensitive feature vector similarity coefficient; carrying out model pre-training on each source domain user by adopting a deep belief network and extracting public mapping knowledge; and finally, the demand response potential of the commercial user can be accurately evaluated through non-parametric method Bayesian ensemble learning model aggregation, and a reference value of the demand response amount of the commercial user is provided for an electric power system operator, so that the scheduling instruction and demand response clearing efficiency is improved, and the labor cost and the computing resource consumption are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of power system automation, and in particular to a response potential evaluation method and system based on temperature-sensitive features and non-parametric transfer learning. Background Art

[0002] Demand response (DR) guides power users to reasonably adjust their energy consumption behavior through price mechanisms or incentive signals. It is an important way for flexible load resources to interact with the power grid flexibly. It can effectively improve the flexible adjustment capability of the power system and play a key role in peak shaving, promoting the consumption of new energy, and maintaining the power balance of the system.

[0003] In the demand-side management of the power trading center, there will be a certain deviation between the actual response volume of some users and their declared capacity, resulting in an uneven distribution of response indicators, thus causing a power gap. Accurate user-side demand response potential assessment helps the power trading center and load aggregators to reasonably regulate demand-side resources and improve the implementation of demand response. When customers accumulate enough historical response data, the data-driven model shows a high evaluation accuracy. However, some users, especially newly signed users, lack historical data, and small-scale training data sets will lead to serious overfitting of the model. Demand response pilot users in various regions usually have sufficient historical response data reserves. Establishing a transfer learning model between pilot users and new participating users is an effective way to solve the above problems. However, when there are significant differences in energy consumption characteristics between users, "negative migration" will occur.

[0004] In addition, traditional transfer learning models can only output certain values ​​rather than probability distribution intervals. For system operators and load aggregators, the confidence interval of demand response potential is more conducive to their evaluation of the upper / lower bounds of user adjustable capabilities and the corresponding uncertainties, thereby formulating more flexible and reliable operation and scheduling strategies. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the traditional transfer learning model can only output a definite value rather than a probability distribution interval, and when there are significant differences in the energy consumption characteristics between users, it will cause "negative transfer"; in order to accurately evaluate the demand response potential of newly signed users and quantify its uncertainty, the purpose of the present invention is to provide a response potential evaluation method and system based on temperature-sensitive features and non-parametric transfer learning, which can accurately evaluate the demand response potential of commercial users and provide power system operators with reference values ​​of commercial users' demand response quantities, thereby improving the efficiency of dispatch instructions and demand response clearing, and reducing labor costs and computing resource consumption.

[0006] The present invention is achieved through the following technical solutions:

[0007] This solution provides a response potential assessment method based on temperature-sensitive features and non-parametric transfer learning, including:

[0008] Obtain the user's historical load data, baseline load data and corresponding temperature data, and calculate the temperature-sensitive load of each user;

[0009] Based on the temperature-sensitive load, a quadratic polynomial fitting is performed to obtain the quadratic regression relationship between the temperature-sensitive load and the temperature change of each user, and a quadratic regression parameter library is constructed based on the quadratic regression relationship;

[0010] Extracting the temperature-sensitive feature vector of each user from the quadratic regression relationship, and calculating the Pearson correlation coefficient of the temperature-sensitive feature vector of each user;

[0011] Obtain the energy consumption pattern of the target user, and screen out source domain users with similar energy consumption patterns to the target user from the quadratic regression parameter library based on the Pearson correlation coefficient, and generate pseudo-label data of demand response for the target user;

[0012] Based on the deep belief network, a demand response potential evaluation model of each source domain user is trained respectively; the demand response potential evaluation model establishes a nonlinear mapping relationship between the characteristic quantity and the user's demand response potential; the characteristic quantity includes: baseline load characteristic quantity, price incentive characteristic quantity and key meteorological characteristic quantity;

[0013] Aggregate the demand response potential evaluation models of multiple source domain users, and derive the conditional probability distribution of the weight vector to evaluate the contribution of different demand response potential evaluation models to the target users; perform weighted calculation on the output values ​​of each demand response potential evaluation model to obtain the demand response evaluation result of the target user.

[0014] A further optimization scheme is that the calculation method of the temperature-sensitive load includes:

[0015] Obtain the benchmark load and temperature-sensitive load of the current user: determine the benchmark reference temperature and benchmark load according to the perceived temperature, the benchmark load includes the benchmark load on weekdays and the benchmark load on weekends; obtain the current user's daily temperature-sensitive load by subtracting the benchmark load at the corresponding time from the power load curve of the target period;

[0016] The actual power load of the current user is obtained by superimposing the benchmark load and the temperature-sensitive load:

[0017]

[0018] in, represents the baseline load of user k at time t, represents the temperature-sensitive load of user k at time t; P t krepresents the actual power load of user k at time t.

[0019] A further optimization scheme is that the quadratic polynomial fitting based on the temperature-sensitive load is performed to obtain the quadratic regression relationship between the temperature-sensitive load and the temperature change of each user; including the method:

[0020] The mapping relationship between the perceived temperature change and the temperature-sensitive load is fitted by a quadratic polynomial function:

[0021]

[0022] Among them, the polynomial coefficients is the temperature-sensitive characteristic coefficient of user k in period t; the temperature-sensitive characteristic vector of user k is They represent the normalized values ​​of the temperature-sensitive load and relative temperature difference of user k in period t on day j; represents the relative temperature difference at time t on day j; Indicates the maximum value of user k's temperature-sensitive load; Indicates the minimum value of user k's temperature-sensitive load; Indicates the maximum value of relative temperature difference; Indicates the minimum value of relative temperature difference.

[0023] A further optimization scheme is that the temperature-sensitive feature vector of each user is extracted from the quadratic regression relationship, and the Pearson correlation coefficient of the temperature-sensitive feature vector of each user is calculated; including a method: using the coefficient of the quadratic polynomial as the temperature-sensitive feature vector of each user, and calculating the Pearson correlation coefficient of the temperature-sensitive feature vector of each user.

[0024] A further optimization scheme is that the energy consumption pattern of the target user is obtained, and source domain users with energy consumption patterns similar to those of the target user are screened out from the quadratic regression parameter library based on the Pearson correlation coefficient, and pseudo-label data of demand response is generated for the target user; including the method:

[0025] Get the temperature-sensitive feature vector of the target user The similarity between the temperature-sensitive feature vectors of each user in the quadratic regression parameter library and the target user is quantitatively evaluated by the Pearson correlation coefficient:

[0026]

[0027] Among them, cov(·) means calculating the covariance of two variables, δ(·) means calculating the standard deviation; ρ k,T represents the similarity between the temperature-sensitive feature vector of user k and the target user, ρ k,T The closer it is to 1, the more similar the temperature-sensitive feature vectors between user k and the target user are; E(·) represents the mathematical expectation; represents the temperature-sensitive feature vectors of the remaining commercial users; represents the i-th element of the temperature-sensitive eigenvector; Represents the square value of each element in the vector; n represents the dimension of the temperature-sensitive feature vector.

[0028] Set a threshold ε, filter out users whose similarity values ​​are greater than the threshold ε as source domain users, and generate pseudo-label data for target users based on the similarity values ​​of source domain users:

[0029]

[0030] in, represents the label data of source domain user k, N S represents the number of source domain users, φS represents the source domain user set, Represents the pseudo-label data of the target user.

[0031] The further optimization scheme is to train the demand response potential evaluation model of each source domain user based on the deep belief network; including the following methods:

[0032] Constructing a deep belief network of each source domain user; the deep belief network is composed of multiple layers of restricted Boltzmann machine layers and fully connected layers, and takes response features as input; the response features include baseline load, body temperature, comfort and price incentives;

[0033] A training set is constructed based on source domain users, and the deep belief network is trained based on the training set to obtain a demand response potential assessment model: first, a multidimensional time series is input into the deep belief network for layer-by-layer unsupervised pre-training to initialize the network parameters of the deep belief network; then, the parameters of the deep belief network are fine-tuned based on the supervised training of the fully connected layer; the training set is Where n represents the number of users in the source domain. represents the response feature set of user i in the source domain; represents the demand response tag data of source domain user i, that is, the load reduction rate of the source domain user when the historical demand response event is triggered; is a real number between 0 and 1.

[0034] A further optimization scheme is that the restricted Boltzmann machine layer consists of a visualization layer V for inputting external high-dimensional time series and a hidden layer h for extracting feature information; the energy function E(V,h) of the restricted Boltzmann machine layer is:

[0035] E(V,h)=-a T Vb T V T W

[0036] Where a and b represent the bias matrices of the visualization layer and the hidden layer respectively, and W represents the weight matrix connecting the visualization layer and the hidden layer;

[0037] The marginal probability distribution of the visualization layer is:

[0038]

[0039] Where p(V,h) represents the joint probability distribution of the visualization layer and the hidden layer; e represents an irrational number;

[0040] Since there is no connection between neurons in the restricted Boltzmann machine layer, the activation actions of the visualization layer and the hidden layer are independent of each other, so the conditional probability distributions of the visualization layer and the hidden layer are:

[0041]

[0042] Among them, m and z represent the number of neurons in the visualization layer and the hidden layer respectively, V i represents the i-th neuron in the visualization layer; h j represents the jth neuron in the hidden layer;

[0043] V i and h j The activation probability is:

[0044]

[0045] Among them, w i,j represents the corresponding element of the i-th row and j-th column in the weight matrix W, G(·) represents the activation function; a i Represents the coefficients of the visualization layer bias matrix; b j represents the coefficients of the hidden layer bias matrix;

[0046]

[0047] Among them, arg max(·) finds the supremum, P(V) is the probability distribution of the visualization layer, and Π(·) is the symbol of multiplication.

[0048] A further optimization scheme is that the gradient calculation process of the weight matrix W includes:

[0049]

[0050] Weight coefficient w i,j The update formula is:

[0051]

[0052] Among them, η represents the learning rate.

[0053] A further optimization scheme is that the non-parametric Bayesian ensemble learning model is used to aggregate the demand response potential evaluation models of multiple source domain users, and the conditional probability distribution of the weight vector is derived to evaluate the contribution of different demand response potential evaluation models to the target user; the output value of each demand response potential evaluation model is weightedly calculated to obtain the demand response evaluation result of the target user; including the method:

[0054] Construct a demand response potential assessment model that is compatible with target users;

[0055] Adaptively assign weights to the output results of the demand response potential assessment model to generate demand response potential assessment results for target users;

[0056] The conditional probability distribution of the weight coefficients of the demand response potential assessment model is derived to generate random weight variables;

[0057] Get the conditional probability distribution of the weight vector corresponding to the random weight variable;

[0058] The target user's demand response evaluation results are obtained based on the conditional probability distribution.

[0059] The present solution also provides a response potential evaluation system based on temperature-sensitive features and non-parametric transfer learning, on which a computer program is stored, characterized in that the computer program is executed by a processor to implement the response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning as described above; the system comprises:

[0060] The acquisition module is used to obtain the user's historical load data, baseline load data and corresponding temperature data, and calculate the temperature-sensitive load of each user;

[0061] A fitting module is used to perform quadratic polynomial fitting based on the temperature-sensitive load to obtain the quadratic regression relationship between the temperature-sensitive load and the temperature change of each user, and to construct a quadratic regression parameter library based on the quadratic regression relationship;

[0062] A first calculation module is used to extract the temperature-sensitive feature vector of each user from the quadratic regression relationship and calculate the Pearson correlation coefficient of the temperature-sensitive feature vector of each user;

[0063] The second calculation module is used to obtain the energy consumption pattern of the target user, and screen out source domain users with similar energy consumption patterns to the target user from the quadratic regression parameter library based on the Pearson correlation coefficient, and generate pseudo-label data of demand response for the target user;

[0064] A training module is used to train a demand response potential evaluation model of each source domain user based on a deep belief network; the demand response potential evaluation model establishes a nonlinear mapping relationship between the feature quantity and the user's demand response potential;

[0065] The evaluation module is used to aggregate the demand response potential evaluation models of multiple source domain users, derive the conditional probability distribution of the weight vector, evaluate the contribution of different demand response potential evaluation models to the target users; perform weighted calculation on the output value of each demand response potential evaluation model to obtain the demand response evaluation result of the target user.

[0066] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0067] The present invention provides a response potential evaluation method and system based on temperature-sensitive features and non-parametric transfer learning. First, a typical temperature-sensitive load characteristic model of commercial users is established, and source domain users that match the energy consumption pattern of potential incremental users are screened out based on the temperature-sensitive feature vector similarity coefficient. A deep belief network is used to pre-train the model of each source domain user and extract common mapping knowledge. Finally, a non-parametric Bayesian ensemble learning model is aggregated to accurately evaluate the demand response potential of commercial users, and provide power system operators with reference values ​​for the demand response amount of commercial users, thereby improving the efficiency of dispatching instructions and demand response clearing, and reducing labor costs and computing resource consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0069] Figure 1 It is a flowchart of the response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning;

[0070] Figure 2 Schematic diagram of the source domain user screening process;

[0071] Figure 3 Schematic diagram of the restricted Boltzmann machine layer structure;

[0072] Figure 4 Schematic diagram of the demand response potential assessment model based on deep belief network;

[0073] Figure 5 Schematic diagram of the transfer learning framework for user demand response potential assessment. DETAILED DESCRIPTION

[0074] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0075] Traditional transfer learning models can only output definite values ​​rather than probability distribution intervals, and when there are significant differences in energy consumption characteristics between users, "negative transfer" will occur; in view of this, the present invention provides the following embodiments to solve the above technical problems.

[0076] Example 1

[0077] This embodiment provides a response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning, such as Figure 1 As shown, including:

[0078] Step 1: Obtain the user's historical load data, baseline load data and corresponding temperature data, and calculate the temperature-sensitive load of each user;

[0079] The calculation method of the temperature sensitive load in this step includes:

[0080] S11, obtaining the benchmark load and temperature-sensitive load of the current user: determining the benchmark reference temperature and benchmark load according to the perceived temperature, wherein the benchmark load includes the benchmark load on weekdays and the benchmark load on weekends; subtracting the benchmark load at the corresponding time from the power load curve of the target period to obtain the temperature-sensitive load of the current user on a daily basis;

[0081] S12, the power load curve of a typical commercial complex is mainly affected by lighting load, plug-in equipment, elevators and other power loads, heating, ventilation and air conditioning (HVAC), among which temperature-sensitive loads such as HVAC are the main source of user demand response adjustment capabilities. In order to describe the temperature-sensitive characteristics of the user, the benchmark load and temperature-sensitive load are superimposed to obtain the actual power load of the current user:

[0082]

[0083] in, represents the baseline load of user k at time t, represents the temperature-sensitive load of user k at time t; P t k represents the actual power load of user k at time t.

[0084] When calculating the benchmark load, the concept of perceived temperature is introduced, and typical days in March and April are selected based on the set perceived temperature threshold. e Obtained according to the following formula:

[0085]

[0086] Among them, T ais the daily average temperature, Rh is the relative humidity, V w represents wind speed; T e A typical day is considered to be between 15°C and 25°C. The base reference temperature, rest day and work day base load are calculated by the following formula:

[0087]

[0088] Among them, ψ week is the set of typical working days in March and April, ψ rest For a typical rest day collection, N w With N r Corresponding to typical working days and rest days, is the workday benchmark load of user k at time t, is the baseline load on rest days, is the reference temperature at time t. The daily temperature-sensitive load can be obtained by subtracting the reference load at the corresponding time from the power load curve in July and August.

[0089]

[0090] in, represents the temperature-sensitive load of user k at time t on day j. The change in perceived temperature is calculated as follows:

[0091]

[0092] in, is the relative temperature difference at time t on the jth day. It should be noted that this study only focuses on the demand response potential during the period of 11:00-21:00, so only the demand response potential of this period, a total of 62 days in July and August, is calculated. The resolution of load data is 15 minutes, and the resolution of body temperature can only reach 1 hour. It is converted into the average value of four consecutive sampling points to represent the temperature-sensitive load data with a resolution of 1h.

[0093] Step 2: Based on the temperature-sensitive load, a quadratic polynomial is fitted to obtain the quadratic regression relationship between the temperature-sensitive load and the temperature change of each user, and a quadratic regression parameter library is constructed based on the quadratic regression relationship; the quadratic regression relationship between the temperature-sensitive load and the temperature change is intended to establish and The display fitting function quantitatively represents the relationship between the degree of change in user energy consumption behavior and the fluctuation of body temperature, and indirectly reflects the amount of power load that can be reduced when users sacrifice a certain degree of body comfort. This solution establishes a quadratic regression relationship between temperature-sensitive load and temperature changes in different time periods, and fits it with a quadratic polynomial function. and The mapping relationship between the temperature change and the temperature-sensitive load is:

[0094]

[0095]

[0096] Among them, the polynomial coefficients is the temperature-sensitive characteristic coefficient of user k in period t; the temperature-sensitive characteristic vector of user k is They represent the normalized values ​​of the temperature-sensitive load and relative temperature difference of user k in period t on day j; represents the relative temperature difference at time t on day j; Indicates the maximum value of user k's temperature-sensitive load; Indicates the minimum value of user k's temperature-sensitive load; Indicates the maximum value of relative temperature difference; Represents the minimum value of relative temperature difference; according to the above, It contains 11 time periods in total.

[0097] Step 3: extract the temperature-sensitive feature vector of each user from the quadratic regression relationship, and calculate the Pearson correlation coefficient of the temperature-sensitive feature vector of each user; specifically, use the coefficient of the quadratic polynomial as the temperature-sensitive feature vector of each user, and calculate the Pearson correlation coefficient of the temperature-sensitive feature vector of each user.

[0098] Step 4: Obtain the energy consumption pattern of the target user, and screen out source domain users with similar energy consumption patterns to the target user from the quadratic regression parameter library based on the Pearson correlation coefficient, and generate pseudo-label data of demand response for the target user; this step specifically includes the following methods:

[0099] Get the temperature-sensitive feature vector of the target user The similarity between the temperature-sensitive feature vectors of each user in the quadratic regression parameter library and the target user is quantitatively evaluated by the Pearson correlation coefficient:

[0100]

[0101] Among them, cov(·) means calculating the covariance of two variables, δ(·) means calculating the standard deviation; ρ k,T represents the similarity between the temperature-sensitive feature vector of user k and the target user, ρ k,T The closer it is to 1, the more similar the temperature-sensitive feature vectors between user k and the target user are; E(·) represents the mathematical expectation; represents the temperature-sensitive feature vectors of the remaining commercial users; represents the i-th element of the temperature-sensitive eigenvector; Represents the square value of each element in the vector; n represents the dimension of the temperature-sensitive feature vector.

[0102] Set a threshold ε, filter out users whose similarity values ​​are greater than the threshold ε as source domain users, and generate pseudo-label data for target users based on the similarity values ​​of source domain users:

[0103]

[0104] in, represents the label data of source domain user k, N S represents the number of users in the source domain, φ S Indicates the source domain user set. Represents the pseudo-label data of the target user.

[0105] The process of source domain user screening and target user pseudo-labeling based on temperature-sensitive feature similarity is as follows: Figure 2 As shown, due to It is not the actual demand response data of the target users, but the weighted average calculated based on the similarity of the source domain user characteristics. Training the demand response potential assessment model for target users using label data will cause large deviations. The auxiliary BPCM evaluates the contribution of different source domain models to the target user demand response potential evaluation process, generates the conditional probability distribution of the weight vector, and outputs the confidence interval of the target user demand response potential to reduce the error of deterministic evaluation.

[0106] Step 5: Based on the deep belief network, a demand response potential evaluation model of each source domain user is trained respectively; the demand response potential evaluation model establishes a nonlinear mapping relationship between the characteristic quantity and the user's demand response potential; the characteristic quantity includes: baseline load characteristic quantity, price incentive characteristic quantity and key meteorological characteristic quantity;

[0107] The demand response potential assessment model aims to establish a nonlinear mapping relationship between the baseline load, price incentive information, key meteorological elements and other characteristic quantities of commercial users and the demand response potential. As a probabilistic generation model, the deep belief network can effectively deal with the volatility and uncertainty of characteristic quantities such as baseline load, humidity, and temperature by probabilistically reconstructing input characteristic information in the form of bidirectional connections. At the same time, it effectively integrates the stable characteristic quantity of price signal with other time-varying high-dimensional characteristic quantities, and realizes the deterministic estimation of the user-side demand response potential by establishing a sequence-sequence implicit fitting function model.

[0108] This step specifically includes the following methods:

[0109] like Figure 3As shown, a deep belief network of each source domain user is constructed; the deep belief network is composed of multiple layers of restricted Boltzmann machine layers and fully connected layers, and takes response features as input; the response features include baseline load, body temperature, comfort and price incentives;

[0110] A training set is constructed based on source domain users, and the deep belief network is trained based on the training set to obtain a demand response potential assessment model: first, a multidimensional time series is input into the deep belief network for layer-by-layer unsupervised pre-training, and the network parameters of the deep belief network are initialized to prevent the gradient from disappearing; then, the parameters of the deep belief network are fine-tuned based on the supervised training full connection layer to further reduce the deviation between the model output and the label data; the training set is Where n represents the number of users in the source domain. represents the response feature set of user i in the source domain; represents the demand response tag data of source domain user i, that is, the load reduction rate of the source domain user when the historical demand response event is triggered; is a real number between 0 and 1.

[0111] like Figure 4 As shown, the restricted Boltzmann machine layer consists of a visualization layer V for inputting external high-dimensional time series and a hidden layer h for extracting feature information; the energy function E(V,h) of the restricted Boltzmann machine layer is:

[0112] E(V,h)=-a T Vb T V T Wh;

[0113] Where a and b represent the bias matrices of the visualization layer and the hidden layer respectively, and W represents the weight matrix connecting the visualization layer and the hidden layer;

[0114] The marginal probability distribution of the visualization layer is:

[0115]

[0116] Where p(V,h) represents the joint probability distribution of the visualization layer and the hidden layer; e represents an irrational number;

[0117] Since there is no connection between neurons in the restricted Boltzmann machine layer, the activation actions of the visualization layer and the hidden layer are independent of each other, so the conditional probability distributions of the visualization layer and the hidden layer are:

[0118]

[0119] Among them, m and z represent the number of neurons in the visualization layer and the hidden layer respectively, V i represents the i-th neuron in the visualization layer; h jrepresents the jth neuron in the hidden layer;

[0120] V i and h j The activation probability is:

[0121]

[0122] Among them, w i,j represents the corresponding element of the i-th row and j-th column in the weight matrix W, G(·) represents the activation function; a i Represents the coefficients of the visualization layer bias matrix; b j represents the coefficients of the hidden layer bias matrix;

[0123]

[0124] Among them, arg max(·) finds the supremum, P(V) is the probability distribution of the visualization layer, and Π(·) is the symbol of multiplication.

[0125] The gradient calculation process of the weight matrix W includes:

[0126]

[0127] Weight coefficient w i,j The update formula is:

[0128]

[0129] Among them, η represents the learning rate.

[0130] Demand response potential assessment is different from load forecasting and new energy output forecasting, which have sufficient data reserves. Even source domain users only have limited labeled data. Deep belief network, as a deep learning algorithm based on probabilistic graph model, can train sequence-to-sequence probability generation model based on limited labeled samples. Its unique two-stage training mode can alleviate the overfitting problem in small sample learning to a certain extent. Therefore, deep belief network is more suitable for demand response potential assessment.

[0131] Step 6: Aggregate the demand response potential evaluation models of multiple source domain users, and derive the conditional probability distribution of the weight vector to evaluate the contribution of different demand response potential evaluation models to the target user; perform weighted calculation on the output values ​​of each demand response potential evaluation model to obtain the demand response evaluation result of the target user. This step specifically includes the following methods:

[0132] S61, construct a demand response potential evaluation model compatible with the target user; the demand response potential evaluation models from different source domains correspond to different nonlinear mapping modes. The essence of sample transfer learning is to aggregate the implicit mapping modes of other commercial users and construct a demand response potential evaluation model compatible with the target user. The specific definition is as follows:

[0133]

[0134] In the formula, The demand response potential assessment results for the target domain. is the input feature of the target domain, Evaluate the model for target users The corresponding parameters of For ensemble learning models The corresponding parameters of Represents the evaluation model of the nth source domain user.

[0135] S61, adaptively assigning weights to the output results of the demand response potential evaluation model to generate a demand response potential evaluation result for the target user;

[0136] Attention mechanism, Stack and other integrated learning models rely on sufficient target domain samples for iterative training to optimize parameters It is not suitable for target users who have not participated in demand response. As a non-parametric ensemble learning model, BPCM only needs a small amount of pseudo-label data to complete the Bayesian reasoning process. The output results are adaptively assigned weights to generate the target user's demand response potential assessment results.

[0137]

[0138] Where N represents the number of commercial users in the source domain. represents the evaluation result output by the i-th source domain evaluation model, is the corresponding weight coefficient, and ε is the residual component.

[0139] S61, deriving the conditional probability distribution of the weight coefficients of the demand response potential assessment model and generating random weight variables;

[0140] The parametric method implements the weight coefficients according to the pre-designed loss function The deterministic estimate of is prone to overfitting when samples are scarce. BPCM does not depend on the parameter Generate fixed weight values, but derive The conditional probability distribution of is used to generate random weight variables, and the contribution of each source domain model to the target business user evaluation model is quantified in the form of uncertainty.

[0141] The weight vector The prior distribution of is given by the Bayesian form:

[0142] S61, obtaining the conditional probability distribution of the weight vector corresponding to the random weight variable;

[0143] The prior distribution is the initial empirical cognition of the weight coefficient, which can be continuously optimized according to the observation results of the labeled samples. Among them, the posterior weight vector ω can be obtained based on Bayesian inference, as shown below:

[0144]

[0145] In the formula Pseudo labels corresponding to target commercial users, Represents the model likelihood.

[0146]

[0147] From the above formula, we can know that the product of the prior and likelihood of the weight vector is proportional to the posterior of the weight vector. Since both the likelihood and the prior distribution satisfy the Gaussian distribution, according to the conjugate distribution theory, the conditional probability distribution of the weight vector can be obtained as follows:

[0148]

[0149] S61, obtaining a demand response evaluation result of the target user according to the conditional probability distribution.

[0150] Where: Once the parameters of the posterior distribution are determined, the output of the model can be evaluated based on each source domain in the test set. Derive the demand response potential of target commercial users Conditional probability distribution.

[0151]

[0152] From the above formula, we can know that The derivation process relies only on a small amount of pseudo-labeled data In order to quantify the uncertainty of the assessment results, The corresponding confidence interval is obtained from the conditional probability distribution, and the interval bandwidth PI under (1-β)×100% confidence level β =[L β ,U β ] is calculated as follows:

[0153]

[0154] Where: L β and Uβ They correspond to the upper and lower bounds of the potential evaluation interval under different confidence levels. BPCM aggregates multiple source domain evaluation models based on a small number of pseudo-label samples and outputs the uncertainty evaluation results of the target business users. The specific process of transfer learning is as follows: Figure 5 shown.

[0155] Example 2

[0156] This embodiment provides a response potential evaluation system based on temperature-sensitive features and non-parametric transfer learning, on which a computer program is stored, and the computer program is executed by a processor to implement the response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning as described in Example 1; the system includes:

[0157] The acquisition module is used to obtain the user's historical load data, baseline load data and corresponding temperature data, and calculate the temperature-sensitive load of each user;

[0158] A fitting module is used to perform quadratic polynomial fitting based on the temperature-sensitive load to obtain the quadratic regression relationship between the temperature-sensitive load and the temperature change of each user, and to construct a quadratic regression parameter library based on the quadratic regression relationship;

[0159] A first calculation module is used to extract the temperature-sensitive feature vector of each user from the quadratic regression relationship and calculate the Pearson correlation coefficient of the temperature-sensitive feature vector of each user;

[0160] The second calculation module is used to obtain the energy consumption pattern of the target user, and screen out source domain users with similar energy consumption patterns to the target user from the quadratic regression parameter library based on the Pearson correlation coefficient, and generate pseudo-label data of demand response for the target user;

[0161] A training module is used to train a demand response potential evaluation model of each source domain user based on a deep belief network; the demand response potential evaluation model establishes a nonlinear mapping relationship between the feature quantity and the user's demand response potential;

[0162] The evaluation module is used to aggregate the demand response potential evaluation models of multiple source domain users, derive the conditional probability distribution of the weight vector, evaluate the contribution of different demand response potential evaluation models to the target users; perform weighted calculation on the output value of each demand response potential evaluation model to obtain the demand response evaluation result of the target user.

[0163] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A response potential evaluation method and system based on temperature-sensitive features and non-parametric transfer learning, characterized in that: include: Obtain the user's historical load data, baseline load data and corresponding temperature data, and calculate the temperature-sensitive load of each user; Based on the temperature-sensitive load, a quadratic polynomial fitting is performed to obtain the quadratic regression relationship between the temperature-sensitive load and the temperature change of each user, and a quadratic regression parameter library is constructed based on the quadratic regression relationship; Extracting the temperature-sensitive feature vector of each user from the quadratic regression relationship, and calculating the Pearson correlation coefficient of the temperature-sensitive feature vector of each user; Obtain the energy consumption pattern of the target user, and screen out source domain users with similar energy consumption patterns to the target user from the quadratic regression parameter library based on the Pearson correlation coefficient, and generate pseudo-label data of demand response for the target user; Based on the deep belief network, a demand response potential evaluation model of each source domain user is trained respectively; the demand response potential evaluation model establishes a nonlinear mapping relationship between the feature quantity and the user's demand response potential; Aggregate the demand response potential evaluation models of multiple source domain users, and derive the conditional probability distribution of the weight vector to evaluate the contribution of different demand response potential evaluation models to the target users; perform weighted calculation on the output values ​​of each demand response potential evaluation model to obtain the demand response evaluation result of the target user.

2. The response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning according to claim 1, characterized in that: The calculation method of the temperature-sensitive load includes: Obtain the benchmark load and temperature-sensitive load of the current user: determine the benchmark reference temperature and benchmark load according to the perceived temperature, the benchmark load includes the benchmark load on weekdays and the benchmark load on weekends; obtain the current user's daily temperature-sensitive load by subtracting the benchmark load at the corresponding time from the power load curve of the target period; The actual power load of the current user is obtained by superimposing the benchmark load and the temperature-sensitive load: in, represents the baseline load of user k at time t, represents the temperature-sensitive load of user k at time t; Represents the actual power load of user k at time t.

3. The response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning according to claim 2, characterized in that: The quadratic polynomial fitting based on the temperature-sensitive load is performed to obtain the quadratic regression relationship between the temperature-sensitive load and the temperature change of each user; Included methods: The mapping relationship between the perceived temperature change and the temperature-sensitive load is fitted by a quadratic polynomial function: Among them, the polynomial coefficients is the temperature-sensitive characteristic coefficient of user k in period t; the temperature-sensitive characteristic vector of user k is They represent the normalized values ​​of the temperature-sensitive load and relative temperature difference of user k in period t on day j; represents the relative temperature difference at time t on day j; Indicates the maximum value of user k's temperature-sensitive load; Indicates the minimum value of user k's temperature-sensitive load; Indicates the maximum value of relative temperature difference; Indicates the minimum value of relative temperature difference.

4. The response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning according to claim 3, characterized in that: The step of extracting the temperature-sensitive feature vector of each user from the quadratic regression relationship and calculating the Pearson correlation coefficient of the temperature-sensitive feature vector of each user; The method includes: using the coefficient of the quadratic polynomial as the temperature-sensitive feature vector of each user, and calculating the Pearson correlation coefficient of the temperature-sensitive feature vector of each user.

5. The response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning according to claim 3, characterized in that: The energy consumption pattern of the target user is obtained, and source domain users with similar energy consumption patterns to the target user are screened out from a quadratic regression parameter library based on the Pearson correlation coefficient, and pseudo-label data of demand response is generated for the target user; including the following method: Get the temperature-sensitive feature vector of the target user The similarity between the temperature-sensitive feature vectors of each user in the quadratic regression parameter library and the target user is quantitatively evaluated by the Pearson correlation coefficient: Among them, cov(·) means calculating the covariance of two variables, δ(·) means calculating the standard deviation; ρ k,T represents the similarity between the temperature-sensitive feature vector of user k and the target user, ρ k,T The closer it is to 1, the more similar the temperature-sensitive feature vectors between user k and the target user are; E(·) represents the mathematical expectation; represents the temperature-sensitive feature vectors of the remaining commercial users; represents the i-th element of the temperature-sensitive eigenvector; Represents the square value of each element in the vector; n represents the dimension of the temperature-sensitive feature vector; Set a threshold ε, filter out users whose similarity values ​​are greater than the threshold ε as source domain users, and generate pseudo-label data for target users based on the similarity values ​​of source domain users: in, represents the label data of source domain user k, N S represents the number of users in the source domain, φ S Indicates the source domain user set. Represents the pseudo-label data of the target user.

6. The response potential assessment method based on temperature-sensitive features and non-parametric transfer learning according to claim 1, characterized in that: Suo Hu trained the demand response potential evaluation model of each source domain user based on the deep belief network; including method: Constructing a deep belief network of each source domain user; the deep belief network is composed of multiple layers of restricted Boltzmann machine layers and fully connected layers, and takes response features as input; the response features include baseline load, body temperature, comfort and price incentives; A training set is constructed based on source domain users, and the deep belief network is trained based on the training set to obtain a demand response potential assessment model: first, a multidimensional time series is input into the deep belief network for layer-by-layer unsupervised pre-training to initialize the network parameters of the deep belief network; then, the parameters of the deep belief network are fine-tuned based on the supervised training of the fully connected layer; the training set is Where n represents the number of users in the source domain. represents the response feature set of user i in the source domain; represents the demand response tag data of source domain user i, that is, the load reduction rate of the source domain user when the historical demand response event is triggered; is a real number between 0 and 1.

7. The response potential assessment method based on temperature-sensitive features and non-parametric transfer learning according to claim 6, characterized in that: The restricted Boltzmann machine layer consists of a visualization layer V that inputs an external high-dimensional time series and a hidden layer h that extracts feature information; the energy function E(V,h) of the restricted Boltzmann machine layer is: E(V,h)=-a T V-b T h-V T Wh Where a and b represent the bias matrices of the visualization layer and the hidden layer respectively, and W represents the weight matrix connecting the visualization layer and the hidden layer; The marginal probability distribution of the visualization layer is: Where p(V,h) represents the joint probability distribution of the visualization layer and the hidden layer; e represents an irrational number; The conditional probability distributions of the visualization layer and the hidden layer are: Among them, m and z represent the number of neurons in the visualization layer and the hidden layer respectively, V i represents the i-th neuron in the visualization layer; h j represents the jth neuron in the hidden layer; V i and h j The activation probability is: Among them, w i,j represents the corresponding element of the i-th row and j-th column in the weight matrix W, G(·) represents the activation function; a i Represents the coefficients of the visualization layer bias matrix; b j represents the coefficients of the hidden layer bias matrix; The training objective of the restricted Boltzmann machine layer is to maximize the expected log probability of the visualization layer in the training set: Among them, argmax(·) finds the supremum, P(V) is the probability distribution of the visualization layer, and Π(·) is the symbol of multiplication.

8. The response potential assessment method based on temperature-sensitive features and non-parametric transfer learning according to claim 7, characterized in that: The gradient calculation process of the weight matrix W includes: Weight coefficient w i,j The update formula is: Among them, η represents the learning rate.

9. The response potential assessment method based on temperature-sensitive features and non-parametric transfer learning according to claim 1, characterized in that: The demand response potential evaluation model of aggregating multiple source domain users is used to derive the conditional probability distribution of the weight vector, and evaluate the contribution of different demand response potential evaluation models to the target user; the output value of each demand response potential evaluation model is weighted to obtain the demand response evaluation result of the target user; including method: Construct a demand response potential assessment model that is compatible with target users; Adaptively assign weights to the output results of the demand response potential assessment model to generate demand response potential assessment results for target users; The conditional probability distribution of the weight coefficients of the demand response potential assessment model is derived to generate random weight variables; Get the conditional probability distribution of the weight vector corresponding to the random weight variable; The target user's demand response evaluation results are obtained based on the conditional probability distribution.

10. A response potential evaluation system based on temperature-sensitive features and non-parametric transfer learning, wherein a computer program is stored thereon, characterized in that: The computer program is executed by a processor to implement the response potential evaluation method based on temperature-sensitive features and non-parametric transfer learning as described in any one of claims 1 to 9; the system includes: The acquisition module is used to obtain the user's historical load data, baseline load data and corresponding temperature data, and calculate the temperature-sensitive load of each user; A fitting module is used to perform quadratic polynomial fitting based on the temperature-sensitive load to obtain the quadratic regression relationship between the temperature-sensitive load and the temperature change of each user, and to construct a quadratic regression parameter library based on the quadratic regression relationship; A first calculation module is used to extract the temperature-sensitive feature vector of each user from the quadratic regression relationship and calculate the Pearson correlation coefficient of the temperature-sensitive feature vector of each user; The second calculation module is used to obtain the energy consumption pattern of the target user, and screen out source domain users with similar energy consumption patterns to the target user from the quadratic regression parameter library based on the Pearson correlation coefficient, and generate pseudo-label data of demand response for the target user; A training module is used to train a demand response potential evaluation model of each source domain user based on a deep belief network; the demand response potential evaluation model establishes a nonlinear mapping relationship between the feature quantity and the user's demand response potential; The evaluation module is used to aggregate the demand response potential evaluation models of multiple source domain users, derive the conditional probability distribution of the weight vector, evaluate the contribution of different demand response potential evaluation models to the target users; perform weighted calculation on the output value of each demand response potential evaluation model to obtain the demand response evaluation result of the target user.

Images