Prediction method of community popularity index based on power load data

Abstract

The present invention discloses a method for predicting a community popularity index based on power load data, which relates to the field of popularity index prediction. Currently, it is difficult to predict the popularity index quickly and accurately. The present invention includes the steps of: obtaining power load data and preprocessing the power load; inputting the preprocessed power load data into a core model to obtain a community population index; outputting the community population index and displaying it. This technical solution utilizes large-scale power data to effectively identify the characteristics of population activities in a region and measure population activity. The popularity index prediction is fast and accurate, with little subjective influence and objective results. Power load data is a by-product of electricity pricing, so predicting the community popularity index based on power load data is low-cost. Obtaining power data is not affected by the environment, and the same power data can be used, which is more conducive to comparative analysis of various regions.

Description

Technical Field

[0001] The present invention relates to the field of popularity index prediction, and in particular to a method for predicting a cell popularity index based on power load data. Background Art

[0002] In recent years, with the development of the economy, the number of traders in commercial areas has increased, and commodities require accurate market information. In addition to manual counting, traditional methods of counting pedestrian flow include infrared sensing counting and gate counting. Among them, the infrared sensing counting method mainly sets up infrared sensors in the places where pedestrians pass, using infrared characteristics to achieve the purpose of pedestrian counting. This method is relatively simple to implement and relatively low in cost, but its statistical accuracy will be greatly affected when multiple pedestrians pass through. Gate counting sets up gates in the places where pedestrians pass, and uses pedestrians to push the gate crossbar to achieve pedestrian flow statistics. This method can relatively accurately count the number of pedestrians, but setting up gates will inevitably affect pedestrian passage, especially when there are many pedestrians, which can easily cause congestion at the gate.

[0003] To address this issue, image acquisition is used to detect the crowd index. However, using images to detect the popularity index has the following drawbacks: image acquisition relies on cameras, which are locally located in certain areas as needed. Camera placement is affected by environmental factors such as buildings, plants, and roads, making it difficult to strictly adhere to uniform settings, which affects the crowd index detection and results in low accuracy. Popularity indices across multiple areas cannot be effectively and accurately compared, limiting their applicability. To mitigate environmental factors, camera locations and angles must be surveyed, planned, and configured upfront, requiring significant time and equipment costs. Furthermore, surveying and planning are subjective, impacting objectivity. Furthermore, the popularity index of a residential area is complex, making it difficult to quickly and accurately determine through image acquisition. Summary of the Invention

[0004] The technical problem to be solved and the technical task proposed by the present invention are to improve and perfect the existing technical solutions and provide a method for predicting the popularity index of a residential area based on power load data, so as to achieve the purpose of objectively and accurately reflecting the popularity index of the residential area. To this end, the present invention adopts the following technical solutions.

[0005] The method for predicting a cell popularity index based on power load data includes the following steps:

[0006] 1) Obtain power load data and pre-process the power load;

[0007] 2) Input the pre-processed power load data into the core model to obtain the community population index;

[0008] 3) Output the community population index and display it.

[0009] As a preferred technical means: in step 2), a popularity index prediction step is included:

[0010] 201) Combining indicators highly correlated with popularity index with implicit representations at the user level; including the average power consumption capacity obtained by fitting a gamma mixture model to the power load distribution and the average user activity obtained by analyzing the average power consumption capacity;

[0011] 202) The user-level latent representation will be aggregated to the community level through mean pooling to obtain the community-level latent representation. This method allows the user latent representation to equally participate in the popularity index prediction;

[0012] 203) The graph encoder GAT summarizes and aggregates the implicit representations of the same timestamp for each community. The graph encoder can effectively help the model extract information;

[0013] 204) The smoothed community latent representation will be passed to the time series autoregressive model GRU for further processing to obtain the final community popularity index prediction.

[0014] As a preferred technical means: in step 1), the steps include:

[0015] 101) Arrange the power load data within the community into 96 load points per user per day and store them separately by date;

[0016] 102) Remove users with long-term missing data from the calculation list;

[0017] 103) Fill in a small amount of null data by taking the average of the nearest neighbors.

[0018] As an optimal technical means: the core model includes a training module, a pre-training module, a probabilistic pre-classification module and a popularity index prediction module; the popularity index prediction is carried out in the popularity index prediction module; the training module is connected with the pre-training module, the probabilistic pre-classification module and the popularity index prediction to obtain the parameters of the popularity index prediction module, thereby obtaining a popularity index prediction model; the pre-training module inputs a series of time series data and outputs the user's latent representation; the probabilistic pre-classification module inputs the user's latent representation and outputs the refined user's latent representation; the popularity index prediction module calculates the popularity index at the specific community level after aggregating the user-level latent representation and outputs it.

[0019] As a preferred technical means: the work of the pre-training module includes the following steps:

[0020] 211) Use predefined multi-layer convolutional neural network The power load data is embedded as local features to perform preliminary processing on the power load data; compared with predictive coding, the encoder g enc The input observation sequence s t-1,t Mapped into a potential information sequence x t , that is, x t =g enc (s t-1,t );

[0021] 212) Using the autoregressive model g ar And through the time series model LSTM, summarize all the information sequences x before time t in the latent space ≤t , and output the context information y at the current time t t =g ar (x ≤t );

[0022] 213) A density ratio model is established to fit the mutual information between the potential information of the observation value and the context information, avoiding the bias of the model training target caused by directly fitting the data.

[0023] As a preferred technical means: the operation of the probability pre-classification module includes the following steps:

[0024] 221: Prompts are used to stimulate the user latent representation of the pre-trained module to extract the refined user latent representation; the probabilistic pre-classification module relies on the auto-regression model g ar , and uses the time series model LSTM to extract semantic information from the time series with prompts; the time series prompt adjustment will refer to the latest results of natural language prompt adjustment to optimize the tiny continuous vector P θ ;

[0025] 222: By further maximizing the soft distance between each pair of heterogeneous data in the same batch, i.e., the latent representations of different users, pre-classification is achieved in the adjustment and the training effect is achieved.

[0026] As a preferred technical means: in step 222), the Softmax between each variable at the same position is calculated; the result is regarded as the probability distribution of the m-classification problem, which is calculated as follows:

[0027]

[0028]

[0029] in, represents the set of semantic representations from m suggestive adjustments for determining user i; To extract the semantic representation at the same position k of each element in the semantic representation set; the probability representation z i,j,k It is represented by semantics Relative to the semantic representation set Y i,k Calculate; y j is a variable;

[0030] After obtaining the probability representation of the user latent representation samples, the KL divergence is calculated to measure the distance between the two probability distributions of the paired user latent representations. The infinity criterion is further introduced for the difference in the probability distribution of the paired user latent representation samples, forming a second loss.

[0031]

[0032]

[0033]

[0034] where Z a ={z a,0 ,z a,1 ,...,z a,m-1} represents the probability representation set Z for extracting the user's implicit representation from m prompt adjustments for determining user a a ; Z a,k ={z a,0,k ,z a,1,k ,...,z a,m-1,k} is the probability representation extracted at the same position k;

[0035] Maximizing the KL divergence directly will lead to a numerical explosion of the KL divergence, because Q(x) is almost zero during the optimization process, causing the divided term in the KL divergence formula to approach infinity; thus, the probability of preventing the expression is close to 1 or close to 0;

[0036] z=max(z,α)

[0037] z=min(z,β)(1)

[0038] Where z is the probability, the lower limit α of z is set to 0.05, and the upper limit β of z is set to 0.95.

[0039] As a preferred technical means: the work of the training module includes the following steps:

[0040] 231) Using power load data to train the pre-training module, the density ratio model of the pre-training model is used to maximize the sample probability;

[0041] 232) Using power load data, the second loss obtained by adding the probability pre-classification module to the pre-training module is trained as a whole; establishing a difference relationship between probability distributions to guide the optimization of user latent representations, thereby further optimizing the ability and confidence of popularity index prediction;

[0042] 233) The popularity index prediction module is trained separately based on the manually labeled relationships between communities and housing price information, so that the popularity index prediction module can determine whether the corresponding community is in a high popularity state, and adjust the popularity index prediction module parameters with the help of the collected information.

[0043] Beneficial effects:

[0044] By implicitly modeling users' daily electricity consumption behavior from electricity load data, and using a label-free probabilistic pre-classification method to optimize user implicit representation, an objective and accurate reflection of the community popularity index can be obtained.

[0045] By leveraging large-scale electricity data, combined with time-series pre-training technology and suggestive adjustment technology, electricity data can be used to reflect the patterns, dynamic changes, and activity levels of population activities. From a micro perspective, individual users' electricity usage behavior indirectly reflects their individual living conditions, trajectories, and quality of life. Combining large-scale user electricity usage and marketing big data can assist in determining the composition of user groups within a region, effectively identifying the characteristics of population activity within the region and measuring population activity. Popularity index predictions are fast and accurate, with minimal subjective influence and objective results. Power load data is a byproduct of electricity pricing, so predicting community popularity indexes based on power load data is low-cost. Obtaining power data is unaffected by the environment, and the same power data can be used, making it more conducive to comparative analysis across regions.

[0046] This technical solution is based on a deep learning framework with time series pre-training and prompt adjustment. Through feature extraction and time series modeling, it predicts the activity level of corresponding points in a community. It quantitatively measures the population density and activity level of a community, tracks changes in this activity level over a given time period, and compares the population activity distribution of other substations to construct a community popularity index prediction model. This model accurately depicts the distribution of population activity in a community over time, while accounting for regional differences. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the design of the probability pre-classification module of the present invention.

[0048] Figure 2 It is a detailed diagram of the probabilistic pre-classification module of the present invention.

[0049] Figure 3 It is a schematic diagram of the present invention.

[0050] Figure 4It is a result display diagram of the present invention. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings.

[0052] As a byproduct of electricity pricing, power companies inevitably record power load data. This data also has minimal relevance to user privacy, making intrusive analysis extremely difficult. Therefore, this technical solution, based on power load data, objectively and accurately reflects the popularity index of a residential area, thereby predicting community popularity, making it a more valuable technical solution.

[0053] like Figure 3 As shown, the method for predicting the cell popularity index based on power load data includes the following steps:

[0054] S1: Obtain power load data and pre-process the power load;

[0055] S2: Input the pre-processed power load data into the core model to obtain the community population index;

[0056] S3: Output the community population index and display it.

[0057] This technical solution is based on a method for detecting the popularity index of a community based on power load data, which achieves the purpose of objectively and accurately reflecting the popularity index of a community. The core algorithm of the present invention, "a method for predicting the popularity index of a community based on power load data", is a deep learning framework based on time series pre-training and prompt adjustment. The algorithm can predict the activity level of corresponding points in a community through methods such as feature extraction and time series modeling. And by constructing a population activity model of residents in the community (substation), the population density and activity of the community are quantitatively measured; and the changes in the activity within a given time period are tracked, and the population activity distribution of other substations is compared to construct a community popularity index prediction model. Under the premise of considering regional differences, the distribution of population activity in the community over a period of time is portrayed.

[0058] Leveraging large-scale electricity data, combined with time-series pre-training and suggestive adjustment techniques, electricity data can be used to reflect the patterns, dynamics, and activity levels of population activity. From a micro perspective, individual electricity usage behavior indirectly reflects their individual lifestyle, trajectory, and quality of life. For example, electricity consumption data can be used to infer the user's morning and evening outing times, usage of high-power appliances, and differences in lifestyle behavior between weekdays and weekends. Furthermore, user payment information can also, to a certain extent, reflect their consumption levels and attitudes. From a macro perspective, the level and patterns of electricity usage in a community can reflect the activity of the population within the area. For example, high overall electricity consumption in a neighborhood likely indicates a high density of users. Time-series characteristics such as the 9-to-5 schedule reflected in overall electricity usage trends can indirectly reflect the behavioral trajectory of most users. Long periods of low electricity consumption by some users often indicate vacant or rented housing. Combining large-scale electricity usage and marketing big data can help determine the composition of user groups within a region, effectively identifying population activity characteristics and measuring population activity within the area.

[0059] The core model framework of this technical solution consists of three main components: a pre-training module that performs preliminary processing of input power load data, a probabilistic pre-classification module, and an objective and accurate popularity index prediction module. The core model executes these three main modules sequentially, using the output of the previous module as the input for the next. Otherwise, the core model relies solely on time series information read from the relevant power company database to generate the target popularity index as output.

[0060] Step S1 is completed by the input module.

[0061] Input Module: The information department authorized by the power company extracts power load data and stores it in the core model. As a byproduct of electricity pricing, power load data is inevitably recorded by power companies. It also has minimal user privacy implications, making intrusive analysis extremely difficult. Therefore, it is extremely valuable to objectively and accurately reflect the popularity index of a community based on power load data. After obtaining data from the power company, the input module's workflow includes the following steps:

[0062] Step 1: Organize the power load data within the community into 96 load points per user per day and store them separately by date.

[0063] Step 2: Remove users with long-term missing data from the calculation list.

[0064] Step 3: Fill in a small amount of null value data by taking the average of the nearest neighbors.

[0065] The core model includes a pre-training module, a probability pre-classification module, a popularity index prediction module and a training module.

[0066] Pre-training module: The input of the pre-training module is a series of time series data and outputs the user's latent representation.

[0067] The working process of the pre-training module includes the following steps:

[0068] Step 1: This module uses a predefined multi-layer convolutional neural network The power load data is embedded as local features to perform preliminary processing on the power load data. Contrastive predictive coding first passes through an encoder g such as a convolutional neural network. enc The input observation sequence s t-1,t Mapped into a potential information sequence x t , that is, x t =g enc (s t-1,t ).

[0069] Step 2: Next, use an autoregressive model g ar , the time series model LSTM is used to summarize all the information sequences x before time t in the latent space ≤t , and output the context information y at the current time t t =g ar (x ≤t ).

[0070] Step 3: The time series pre-training model in this module does not directly use the generative model to predict future observations and calculate the corresponding gaps. Instead, this method establishes a density ratio model to fit the latent information x of the observations. t+k and context information y t The mutual information between them is given by the following formula. This method avoids the bias of the model training target caused by directly fitting the data.

[0071]

[0072]

[0073] The convolutional neural network of this module is a 3-layer one-dimensional convolutional network, and the autoregressive model is a 2-layer LSTM.

[0074] Probabilistic Pre-Classification Module: This module takes as input the user's latent representation and outputs a highly refined version of that representation. This module guides the optimization of the user's latent representation by establishing a differential relationship between probability distributions, thereby further improving the ability and confidence of popularity index prediction.

[0075] like Figure 1As shown, the probabilistic pre-classification module uses a prompt adjustment strategy to input the prompt words into the autoregressive model g in parallel on the predicted version of the context information in each prompt adjustment. ar These predictions are not immediately aggregated or concatenated, but instead retained for a new form of pre-classification model. A multi-view mechanism is introduced, allowing multiple prompt adjustments to jointly focus on information generated by the model from prompts with different meanings. This mechanism enables prompt adjustments to express a wider range of semantic information.

[0076] like Figure 2 As shown in the figure, in the probabilistic pre-classification model, the representations at the same position from different suggestive adjustments are assumed to represent different aspects of the morpheme concept. The probabilistic pre-classification model for m prompts first calculates the softmax between the variables at the same position. This result is considered the probability distribution of the m-classification problem. After obtaining the probabilistic representation of the data samples, the model calculates the KL divergence to measure the distance between the two probability distributions of paired data samples. The model further introduces an infinity criterion for the difference in the probability distributions of paired data samples, which forms the second training loss of the probabilistic pre-classification module.

[0077] The working process of the probabilistic pre-classification module includes the following steps:

[0078] Step 1: This module stimulates the user latent representation of the pre-trained module through prompts to extract the refined user latent representation. The probabilistic pre-classification module relies on the original auto-regression model g ar , the time series model LSTM is used to extract semantic information from the time series with prompts. The time series prompt adjustment will refer to the latest results of natural language prompt adjustment to optimize the small continuous vector P θ This tiny continuous vector is similar to the word "summary" in natural language.

[0079] Step 2: This module further maximizes the soft distance between each pair of heterogeneous data within the same batch, i.e., between different users' latent representations, to achieve pre-classification during adjustment and maintain the training effect. The probabilistic pre-classification model for m (m is a constant) prompts first calculates the softmax between each variable at the same position. This result is considered the probability distribution of the m-classification problem. This calculation process is as follows.

[0080]

[0081]

[0082] in, represents the set of semantic representations from m suggestive adjustments; This means that the semantic representation will be extracted at the same position k of each element in the set of semantic representations. In this equation, the probability representation z i,j,k It is represented by semantics Relative to the semantic representation set Y i,k Calculated.

[0083] After obtaining the probabilistic representations of user latent representation samples, the model calculates the KL divergence to measure the distance between the two probability distributions of paired user latent representations. Similar to the approach used in suggestive adjustment, the model further introduces an infinity criterion for the difference in the probability distributions of paired user latent representation samples, forming the following second loss.

[0084]

[0085]

[0086]

[0087] where Z a ={z a,0 ,z a,1 ,...,z a,m-1} represents the probability representation set Z of extracting the user's latent representation sample a from m prompt adjustments a ; Z a,k ={z a,0,k ,z a,1,k ,...,z a,m-1,k} means that the probability representation will be extracted at the same position k.

[0088] Maximizing the KL divergence directly will lead to a numerical explosion of the KL divergence, because Q(x) is almost zero during the optimization process, causing the divided term in the KL divergence formula to approach infinity. This prevents z from approaching 1 or 0.

[0089] z=max(z,α)

[0090] z=min(z,β) (5)

[0091] Based on experience, the lower limit α of z is set to 0.05, and the upper limit β of z is set to 0.95.

[0092] The convolutional neural network of this module is a 3-layer one-dimensional convolutional network, the autoregressive model is a 2-layer LSTM, and the prompt settings are fixed learnable parameters.

[0093] Popularity Index Prediction Module: This module aggregates user-level latent representations to calculate the popularity index at the community level. The module directly outputs the popularity index for the corresponding community. The module consists of a graph encoder (GAT) and an autoregressive model (GRU), followed by a linear layer and a sigmoid function.

[0094] The popularity index prediction module works in the following steps:

[0095] Step 1: The popularity index prediction module combines several metrics highly correlated with popularity index with user-level latent representations. These include an average power consumption capacity obtained by fitting a gamma mixture model to the power load distribution and an average user activity level obtained by analyzing the average power consumption capacity.

[0096] Step 2: The user-level latent representation is aggregated to the community level through mean pooling to obtain the community-level latent representation. This method allows the user latent representation to participate equally in the popularity index prediction.

[0097] Step 3: The graph encoder GAT summarizes and aggregates (smooths) the latent representations of the same timestamp for each community. The graph encoder can effectively help the model extract information.

[0098] Step 4: The smoothed community latent representation will be passed to the time series autoregressive model GRU for further processing to obtain the final community popularity index prediction.

[0099] The graph encoder GAT of the popularity index prediction module uses GATv2Conv, which has 128 hidden dimensions and is implemented by PyTorch Geometric. The autoregressive model is a 2-layer GRU with 128 hidden dimensions.

[0100] Training module: The CyclicLR optimizer is used during training. The popularity index prediction module uses mean absolute error (L1-loss) to calculate the loss. The model is divided into three stages of training to adjust the model parameters. The training module process includes the following steps:

[0101] Step 1: In the first stage, the pre-training module is trained using power load data. The sample probability is maximized by the density ratio of the pre-trained model.

[0102] Step 2: In the second phase, the power load data is used to add the second loss introduced in the probabilistic pre-classification module to the original pre-training module, and then the two modules are trained as a whole. The difference relationship between the probability distribution is established to guide the optimization of the user's latent representation, thereby further improving the ability and confidence of popularity index prediction.

[0103] Step 3: Finally, the popularity index prediction module is trained separately based on the manually labeled relationships between communities and housing price information, so that the popularity index prediction model can determine whether the corresponding community is in a high popularity state, and use the collected information to adjust the popularity index prediction module parameters.

[0104] The training process is all based on the gradient descent algorithm.

[0105] Output and visualization module: The popularity index prediction module generates the popularity index, and stores the user implicit representations calculated in other modules, distributes the popularity index to the appropriate granularity and generates corresponding visualization results. Figure 4 shown.

[0106] Experimental verification

[0107] 1. Data Description

[0108] In order to verify the superiority of the method, experiments were carried out on a residential electricity load dataset.

[0109] Residential Electricity Load Data: This dataset, provided by China State Grid Zhejiang Electric Power Co., Ltd. and Hangzhou Huayun Technology Co., Ltd., contains 96 daily load data points from over 20,000 electricity meters in Xihu and Jianggan Districts of Hangzhou, Zhejiang Province, China, between July 1, 2019, and July 1, 2020. Each meter corresponds to a community, and its latitude and longitude can be retrieved from the corresponding structured address. For privacy reasons, each meter is anonymized; sensitive information such as the house number is initially hidden. Generating 96 daily load data points means that the electricity load is sampled and generated every 15 minutes. Using this dataset, we will learn a model that derives from electricity load data to community-level social variables, such as housing prices, which are highly correlated with community economic conditions.

[0110] 2. Popularity Index Fitting Experiment

[0111] The proposed multi-view time series suggestive adjustment is compared with different pre-trained models and adjustment strategies: (1) contrastive predictive coding and three time-scale pooling methods; (2) TS-TCC: a contrastive learning framework using data augmentation; (3) SimCLR; and (4) a cell popularity index prediction model based on power load data. The three time-scale pooling methods are global average aggregation, retaining the last item in the sequence, and concatenating all hidden layer outputs. In the current task experiment, the cell popularity index prediction model based on power load data outperforms all other methods.

[0112] Table 1. Average performance of popularity index fitting experiments.

[0113]

[0114] The community popularity index prediction method based on power load data shown above is a specific embodiment of the present invention, which has reflected the essential characteristics and progress of the present invention. According to actual usage needs and under the guidance of the present invention, equivalent modifications in shape, structure, etc. can be made to it, which are all within the scope of protection of this scheme.

Claims

1. A method for predicting community popularity index based on power load data, characterized by The following steps are involved: 1) Obtain power load data and pre-process the power load; 2) Input the pre-processed power load data into the core model to obtain the community population index; 3) Output and display the community population index; In step 2), the popularity index prediction step is included: 201) Combining metrics highly correlated with popularity index with implicit representations at the user level; including average power consumption obtained by fitting a gamma mixture model to the power load distribution and average user activity obtained by analyzing the average power consumption; 202) The user-level latent representation is aggregated to the community level through mean pooling to obtain the community-level latent representation. This method allows the user-level latent representation to participate equally in the popularity index prediction; 203) The graph encoder GAT aggregates and summarizes the implicit representations of the same timestamp for each community. The graph encoder can effectively help the model extract information. 204) The smoothed community latent representation will be passed to the time series autoregressive model GRU for further processing to obtain the final community popularity index prediction; The core model includes a training module, a pre-training module, a probabilistic pre-classification module and a popularity index prediction module; the popularity index prediction is carried out in the popularity index prediction module; the training module is connected with the pre-training module, the probabilistic pre-classification module and the popularity index prediction to obtain the parameters of the popularity index prediction module, thereby obtaining a popularity index prediction model; the pre-training module inputs a series of time series data and outputs the user's latent representation; the probabilistic pre-classification module inputs the user's latent representation and outputs the refined user's latent representation; the popularity index prediction module calculates the popularity index at the specific community level after aggregating the user-level latent representation and outputs it.

2. The method for predicting a cell popularity index based on power load data according to claim 1, characterized in that: In step 1), the steps include: 101) Organize the community-wide power load data into 96 load points per user per day and store them by date; 102) Remove users with long-term missing data from the calculation list; 103) Fill in a small amount of null data by taking the average of the nearest neighbors.

3. The method for predicting a cell popularity index based on power load data according to claim 1, characterized in that: The work of the pre-training module includes the following steps: 211) Use predefined multi-layer convolutional neural networks The power load data is embedded as local features to perform preliminary processing on the power load data; compared with predictive coding, the encoder of the convolutional neural network is used The input observation sequence Mapping to potential information sequence ,Right now ; 212) Using an autoregressive model And through the time series model LSTM, summarize the information sequence before all t moments in the latent space , and output the context information at the current time t ; 213) A density ratio model is established to fit the mutual information between the latent information of the observations and the contextual information, avoiding the bias of the model training objectives caused by directly fitting the data.

4. The method for predicting a cell popularity index based on power load data according to claim 3, characterized in that: The operation of the probabilistic pre-classification module includes the following steps: 221) Prompts are used to stimulate the user latent representation of the pre-trained module to extract the refined user latent representation; the probabilistic pre-classification module relies on the auto-regression model , and uses the time series model LSTM to extract semantic information from the time series with prompts; the time series prompt adjustment will refer to the latest results of natural language prompt adjustment to optimize small continuous vectors ; 222) By further maximizing the soft distance between each pair of heterogeneous data in the same batch, i.e., the latent representations of different users, pre-classification is achieved in the adjustment and the training effect is achieved.

5. The method for predicting a cell popularity index based on power load data according to claim 4, characterized in that: In step 222), the Softmax between the variables at the same position is calculated; the result is regarded as the probability distribution of the m-classification problem, which is calculated as follows: in, represents the set of semantic representations from m suggestive adjustments for a certain user i; To extract the semantic representation at the same position k of each element in the semantic representation set; probability representation It is represented by semantics Relative to the semantic representation set calculate; is a variable; After obtaining the probability representation of the user latent representation samples, the KL divergence is calculated to measure the distance between the two probability distributions of the paired user latent representations. The infinity criterion is further introduced for the difference in the probability distribution of the paired user latent representation samples, forming a second loss. in Represents the probability representation set for extracting user hidden from m prompt adjustments for determining user a ; To extract the probability representation at the same position k; Directly maximizing the KL divergence will lead to an explosion in the value of the KL divergence because During the optimization process, it is almost zero, which causes the divided term in the KL divergence formula to approach infinity; therefore, the blocking probability is close to 1 or close to 0; Where z is the probability, and the lower limit of z is Set to 0.05, the upper limit of z Set to 0.

95.

6. The method for predicting a cell popularity index based on power load data according to claim 5, characterized in that: The training module includes the following steps: 231) Use power load data to train the pre-training module and maximize the sample probability through the density ratio model of the pre-training model; 232) Using power load data, we add the second loss obtained from the probabilistic pre-classification module to the pre-training module, and train both modules as a whole. We establish a differential relationship between probability distributions to guide the optimization of user latent representations, thereby further improving the ability and confidence of popularity index prediction. 233) The popularity index prediction module is trained separately based on the manually labeled relationships between communities and housing price information, so that the popularity index prediction module can determine whether the corresponding community is in a high popularity state, and use the collected information to adjust the popularity index prediction module parameters.

Images