A method for evaluating the potential of grid demand response in near-zero energy buildings

Abstract

The present invention relates to a method for evaluating the demand response potential of a near-zero energy building power grid. A convolutional neural network (CNN) is used to extract local features from multi-source data, and a variational autoencoder (VAE) is used to perform feature dimension reduction and L1 regularization feature selection, so that the feature data set is more representative and highly interpretable. A long short-term memory (LSTM) network is introduced to effectively capture the long-term temporal dependency in power demand. At the same time, a self-attention mechanism is introduced into the LSTM layer to dynamically adjust the model's attention to features at different time steps, thereby more accurately reflecting the details of demand changes. An adaptive reinforcement learning model based on a deep Q network (DQN) is used to perform self-adjustment and optimization according to real-time data, dynamically adjust the demand response strategy, and improve the flexibility and effect of demand response. Not only is the training stability improved through the experience replay mechanism, but the generalization ability of the model under different situations can also be guaranteed through the target network update mechanism.

Description

Technical Field

[0001] The invention belongs to the technical field of power grid demand response, and in particular to a method for evaluating the potential of power grid demand response in a near-zero energy consumption building. Background Art

[0002] As the world grapples with climate change and global warming, industries are seeking to minimize carbon emissions and turn to renewable energy sources; this is especially true for the construction industry, which is often accused of having a huge impact on the environment; this is because a large amount of energy, natural resources and raw materials are used, a large amount of waste is generated, and carbon that is harmful to the earth is emitted during the construction, use, renovation and even demolition of buildings; however, the construction sector can significantly reduce energy use by incorporating energy-saving strategies into the design, construction and operation of new buildings, and by retrofitting existing buildings to improve efficiency; at the same time, the coordination with near-zero energy buildings includes grid demand response. Demand response, as an important part of the implementation of electricity demand side management, refers to the market participation behavior that guides and encourages electricity users to actively change their original electricity consumption patterns through market price signals such as time-of-use electricity prices or incentive mechanisms such as financial subsidies, so as to promote the balance of electricity supply and demand and ensure the stable operation of the power grid. The combination of near-zero energy buildings and grid demand response further reduces energy use.

[0003] For example, the patent with publication number CN115829119A discloses a method for evaluating the demand-side response potential and load transfer of a power grid based on neural networks and reinforcement learning, which includes the following steps: using an improved k-means clustering algorithm to cluster massive amounts of customer electricity consumption data to generate electricity consumption behavior labels, and clustering customer data into three categories: peak-free, single-peak, and multi-peak; building and training a multi-layer perceptron (MLP) neural network model, using the customer's 96-point workday load data and electricity consumption behavior labels as input, and mining the implicit mapping relationship between the input data and the customer's willingness and potential to participate in demand response; using deep reinforcement learning to solve the constructed load transfer model, and outputting the optimal time-of-use electricity price formulation recommendation; the application fully mines and utilizes the characteristics of customer electricity consumption data, and then predicts the customer's willingness and potential to participate in demand response; improving the peak-shaving and valley-filling capabilities of the power grid, alleviating the contradiction between electricity supply and demand, and ensuring the stable operation of the power grid.

[0004] Grid demand response refers to a mechanism in which electricity users adjust their electricity usage behavior according to grid demand and price signals to help the grid balance supply and demand and improve stability. Under this mechanism, the nonlinear and time-varying characteristics of electricity demand fluctuations become key factors in prediction and response. In terms of nonlinear characteristics, electricity demand is affected by a variety of highly nonlinear factors, including temperature, season, economic activities and user behavior. In terms of time-varying characteristics, electricity demand exhibits different patterns in different time periods and has significant time-varying properties. Both short-term fluctuations and long-term trend changes need to be accurately captured. Therefore, a method for evaluating the potential of grid demand response in near-zero energy buildings is urgently needed to solve such problems. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and propose a method for evaluating the demand response potential of a near-zero energy building power grid, which can solve the problem in the prior art that it is unable to effectively capture the power demand fluctuations with nonlinear and time-varying characteristics.

[0006] The present invention solves the technical problem by adopting the following technical solutions:

[0007] A method for evaluating the potential of grid demand response of a near-zero energy building comprises the following steps:

[0008] Step 1. Grid demand data collection: collect power load data, weather data, economic activity data, and user behavior data; process missing values, outliers, and noise data, and standardize and normalize the collected data;

[0009] Step 2. Extract the features of grid demand data. Use the convolutional neural network (CNN) to extract local features from multi-source data, extract temperature change patterns from weather data, and extract load fluctuation features from electricity consumption data. Slice the time series data to extract key time series features, including the peak and valley time periods of daily electricity consumption.

[0010] Use variational autoencoder VAE to reduce feature dimensionality, remove redundant information and retain important features, and select the most representative features through L1 regularization method;

[0011] Step 3. Multi-level deep learning model construction: construct a multi-level model combining convolutional neural network (CNN) and long short-term memory (LSTM) networks. Convolutional neural network (CNN) is used to extract local features, and long short-term memory (LSTM) networks are used to capture long-term temporal dependencies. Introduce a self-attention mechanism in the LSTM layer to dynamically adjust the model's attention to features at different time steps.

[0012] Train a multi-level deep learning model using the feature dataset extracted and selected in step 2; and use a cross-validation method to evaluate the model performance;

[0013] Step 4. Introduce adaptive reinforcement learning to obtain real-time power demand and environmental data from the smart grid and update the data set in real time; build an adaptive reinforcement learning model based on the deep Q network DQN to dynamically adjust the demand response strategy; the model self-adjusts and optimizes according to real-time data; and use the prediction results of the multi-level deep learning model as the input of the reinforcement learning model;

[0014] Step 5. Evaluate the demand response potential. Use the trained multi-level deep learning model and adaptive reinforcement learning model to predict electricity demand and identify potential demand response opportunities. Combine historical data and real-time data to analyze the impact of various factors on electricity demand and evaluate the demand response potential under different circumstances. Select key evaluation indicators, including prediction accuracy, response time, and user engagement, to comprehensively evaluate the potential for demand response. Identify the optimal demand response plan by comparing the effects of different strategies.

[0015] Moreover, the convolutional neural network CNN in step 2 includes:

[0016] Input layer: input preprocessed multi-source data;

[0017] Convolutional layer: deploy multiple convolutional layers and use convolution kernels of different sizes to extract local features;

[0018] Activation function: Apply a non-linear activation function after the convolutional layer;

[0019] Pooling layer: Add pooling layer as needed to reduce feature dimensions;

[0020] Fully connected layer: Map the pooled features to the fully connected layer to obtain the feature vector; then train the convolutional neural network CNN;

[0021] Extract temperature change patterns from weather data; extract load fluctuation characteristics from electricity consumption data.

[0022] Moreover, the specific implementation method of slicing the time series data and extracting key time series features in step 2 is:

[0023] Slice the power load data at fixed time intervals to form time series segments;

[0024] Extract key time series features, including daily peak power consumption, identifying the time period with the highest daily power consumption; daily trough power consumption, identifying the time period with the lowest daily power consumption;

[0025] Analyze the periodicity and trend characteristics of electricity consumption data, including the difference in electricity consumption patterns between weekdays and weekends;

[0026] The variational autoencoder VAE is used for feature dimensionality reduction. The variational autoencoder VAE includes:

[0027] Input layer: input high-dimensional features extracted from multi-source data;

[0028] Encoder: maps high-dimensional input features to a low-dimensional latent space;

[0029] Decoder: reconstructs the original features from the low-dimensional latent space;

[0030] Loss function: including reconstruction loss and KL divergence;

[0031] Train the variational autoencoder VAE to map high-dimensional features to a low-dimensional latent space, remove redundant information, and retain important features.

[0032] Moreover, the specific implementation method of using the L1 regularization method to select the most representative features in step 2 is:

[0033] Linear regression, logistic regression and Lasso regression are used to build models with L1 regularization, and L1 regularization terms are added to the loss function;

[0034] Input the reduced-dimensional features into the model, and train and evaluate the model with L1 regularization;

[0035] Through L1 regularization, a sparse feature set is selected, and features with weights close to zero are removed; the features that contribute most to the model prediction effect are retained.

[0036] Moreover, the multi-level model combining the convolutional neural network CNN and the long short-term memory network LSTM constructed in step 3 includes:

[0037] Input layer: input preprocessed multi-source data ,in is the number of samples, is the number of time steps, is the number of features;

[0038] Convolutional layer CNN: Use multiple convolutional layers to extract local features: ,in It is The output feature map of the layer, It is Tier convolution kernels, It is The input feature map of the layer, It is The bias of the layer, is the ReLU activation function, Represents the convolution operation;

[0039] Pooling layer: Apply pooling operation to reduce feature dimension. , It is The pooled output feature map of the layer, is the pooling function;

[0040] Expand the pooled feature map into a time step feature sequence ,in It is the output feature dimension of the pooling layer;

[0041] Long Short-Term Memory Network LSTM layer:

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] in , , are the activation vectors of the input gate, forget gate, and output gate, respectively. is the memory unit state; is the hidden state, is the time step The input feature vector is , , , are the weight matrices of the input gate, forget gate, output gate, and memory unit, respectively. , , , are the weight matrices from the previous hidden state to the input gate, forget gate, output gate, and memory unit, respectively. , , , are the bias vectors of the input gate, forget gate, output gate and memory unit respectively, is the Sigmoid activation function, is the hyperbolic tangent activation function, Represents element-wise multiplication.

[0048] Moreover, the specific implementation method of introducing the self-attention mechanism in the long short-term memory network LSTM layer in step 3 is:

[0049] Calculate the attention weights between time steps: ,in is the time step For time step The attention score, is the weight vector, is the transpose of the weight vector, and are the weight matrices of the hidden states of the current time step and all time steps, respectively. is the bias vector;

[0050] Calculate the normalized attention weights: ,in is the time step For time step The attention weight of

[0051] Calculate the attention-weighted context vector: , is the time step The attention-weighted context vector of ;

[0052] Combine the attention-weighted context vector with the hidden state of the LSTM, , is the time step The final output of

[0053] and are the weight matrices of the context vector and the hidden state, respectively; the feature dataset extracted and selected in step 2 is used to train a multi-level model combining a convolutional neural network CNN and a long short-term memory network LSTM, and a self-attention mechanism is introduced.

[0054] Moreover, the adaptive reinforcement learning model based on the deep Q network DQN constructed in step 4 is:

[0055] Define states, actions, and rewards:

[0056] state : Current power demand and environmental data, including power consumption, temperature, and humidity;

[0057] action : Demand response strategies, including increasing and decreasing load;

[0058] award : Grid stability and economic benefits after demand response, including load balancing and cost savings;

[0059] Q value update formula: ,in is the state action value function, which means that in the state Take Action Expected rewards; is the learning rate, It is a timely reward, reflecting the effectiveness of the current demand response strategy. is the discount factor, which indicates the influence of future rewards on current decisions. The next state The maximum Q value of

[0060] Deep Q network structure:

[0061] Input layer: input current state ;

[0062] Hidden layer: multiple fully connected layers, using the activation function ReLU to improve the nonlinear representation ability of the model;

[0063] Output layer: outputs the Q value of each action;

[0064] The training process is:

[0065] Experience Replay: Storing State Transition Samples , randomly select small batches of samples for training;

[0066] Target network: Use the target network to improve training stability and update the target network parameters every fixed number of steps.

[0067] Moreover, the specific implementation method of predicting power demand and identifying potential demand response opportunities in step 5 is:

[0068] Use the trained multi-layer deep learning model to predict future electricity demand;

[0069] The forecast results include short-, medium- and long-term electricity demand;

[0070] Identify peak load periods and off-peak load periods by analyzing forecasted electricity demand data;

[0071] During peak load periods, identify time windows where demand response can be implemented and assess the potential of demand response.

[0072] Moreover, the specific implementation method of analyzing the impact of various factors on power demand in step 5 is:

[0073] Utilize historical electricity demand data to analyze the impact of various factors on electricity demand;

[0074] Use statistical methods to quantify the impact of various factors on electricity demand;

[0075] Utilize real-time data to analyze the impact of current environment and demand changes on electricity demand;

[0076] Compare real-time data with historical data, identify abnormal changes, and adjust demand response strategies;

[0077] Combine historical and real-time data to conduct comprehensive analysis and evaluate the demand response potential under different scenarios.

[0078] The advantages and positive effects of the present invention are:

[0079] 1. The present invention extracts local features from multi-source data through convolutional neural network (CNN), combines variational autoencoder (VAE) to perform feature dimension reduction and L1 regularization to select features, making the feature data set more representative and highly interpretable.

[0080] 2. The present invention introduces the long short-term memory network LSTM, which effectively captures the long-term temporal dependency in electricity demand; at the same time, the self-attention mechanism is introduced in the LSTM layer, which can dynamically adjust the model's attention to features of different time steps, thereby more accurately reflecting the details of demand changes.

[0081] 3. The present invention optimizes the demand response strategy that relies on static rules. Through an adaptive reinforcement learning model based on the deep Q network DQN, it performs self-adjustment and optimization according to real-time data, dynamically adjusts the demand response strategy, and improves the flexibility and effectiveness of demand response. Not only can the training stability be improved through the experience replay mechanism, but the generalization ability of the model in different situations can also be guaranteed through the target network update mechanism. The present invention solves the problem in the prior art that it is impossible to effectively capture the power demand fluctuations with nonlinear and time-varying characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0083] The present invention is further described in detail below with reference to the accompanying drawings.

[0084] The present invention provides a method for evaluating the demand response potential of a near-zero energy building power grid. Figure 1 As shown, including:

[0085] Step 1. Grid demand data collection: collect power load data (historical power consumption records), weather data (temperature, humidity, wind speed, etc.), economic activity data (production activities, commercial activities, etc.) and user behavior data (power consumption habits, demand changes, etc.); process missing values, outliers and noise data, and standardize and normalize the collected data;

[0086] Step 2. Extract the features of grid demand data. Use the convolutional neural network (CNN) to extract local features from multi-source data, extract temperature change patterns from weather data, and extract load fluctuation features from electricity consumption data. Slice the time series data to extract key time series features, including the peak and valley time periods of daily electricity consumption.

[0087] Use variational autoencoder VAE to reduce feature dimensionality, remove redundant information and retain important features, and select the most representative features through L1 regularization method; construct convolutional neural network CNN, and then perform convolutional neural network CNN training and feature extraction;

[0088] The convolutional neural network CNN architecture includes:

[0089] Input layer: input preprocessed multi-source data (weather data, electricity consumption data);

[0090] Convolutional layer: deploy multiple convolutional layers and use convolution kernels of different sizes (3x3, 5x5) to extract local features;

[0091] Activation function: Apply a nonlinear activation function (ReLU) after the convolutional layer to enhance the nonlinear expression ability of the model;

[0092] Pooling layer: Add a pooling layer (maximum pooling or average pooling) as needed to reduce feature dimensions and enhance feature robustness;

[0093] Fully connected layer: Map the pooled features to the fully connected layer to obtain the feature vector; then train the convolutional neural network CNN;

[0094] Extract temperature variation patterns from weather data; extract load fluctuation characteristics from electricity consumption data;

[0095] Slice the time series data and extract key time series features:

[0096] Slice the power load data at fixed time intervals (hours, days, weeks) to form time series segments;

[0097] Extract key time series features, including daily peak power consumption, identifying the time period with the highest daily power consumption; daily trough power consumption, identifying the time period with the lowest daily power consumption;

[0098] Analyze the periodicity and trend characteristics of electricity consumption data, including the difference in electricity consumption patterns between weekdays and weekends;

[0099] The variational autoencoder VAE is used for feature dimensionality reduction. The variational autoencoder VAE includes:

[0100] Input layer: input high-dimensional features extracted from multi-source data;

[0101] Encoder: maps high-dimensional input features to a low-dimensional latent space;

[0102] Decoder: reconstructs the original features from the low-dimensional latent space;

[0103] Loss function: including reconstruction loss and KL divergence, which enables the model to remove redundant information while retaining important features;

[0104] Train the variational autoencoder VAE to map high-dimensional features to a low-dimensional latent space, remove redundant information, and retain important features;

[0105] Use L1 regularization method to select the most representative features:

[0106] Linear regression, logistic regression, and Lasso regression are used to build models with L1 regularization, and L1 regularization terms are added to the loss function;

[0107] Input the reduced-dimensional features into the model, and train and evaluate the model with L1 regularization;

[0108] Through L1 regularization, a sparse feature set is selected, and features with weights close to zero are removed; the features that contribute most to the model prediction effect are retained.

[0109] Step 2 effectively extracts key features from multi-source data, performs feature dimensionality reduction and selection, uses convolutional neural network (CNN) to extract local features, variational autoencoder (VAE) to perform feature dimensionality reduction, and L1 regularization to select the most representative features, making the feature data set used for demand response evaluation highly representative and interpretable.

[0110] Step 3. Multi-level deep learning model construction: construct a multi-level model combining convolutional neural network (CNN) and long short-term memory (LSTM). Convolutional neural network (CNN) is used to extract local features, and long short-term memory (LSTM) is used to capture long-term temporal dependencies. Introduce a self-attention mechanism in the LSTM layer to dynamically adjust the model's attention to features at different time steps.

[0111] Use the feature dataset extracted and selected in step 2 to train a multi-level deep learning model and optimize the parameters so that the model can accurately capture the nonlinear and time-varying characteristics of electricity demand;

[0112] Use cross-validation method to evaluate model performance, prevent overfitting, and ensure the generalization ability of the model on different data sets;

[0113] The multi-level model architecture combining CNN and LSTM is:

[0114] Input layer: input preprocessed multi-source data ,in is the number of samples, is the number of time steps, is the number of features;

[0115] Convolutional layer CNN: Use multiple convolutional layers to extract local features: ,in It is The output feature map of the layer, It is Tier convolution kernels, It is The input feature map of the layer, It is The bias of the layer, is the ReLU activation function, Represents the convolution operation;

[0116] Pooling layer: Apply pooling operation to reduce feature dimension. , It is The pooled output feature map of the layer, is the pooling function;

[0117] Expand the pooled feature map into a time step feature sequence ,in It is the output feature dimension of the pooling layer;

[0118] Long Short-Term Memory Network LSTM layer:

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] in , , are the activation vectors of the input gate, forget gate, and output gate, respectively. is the memory unit state; is the hidden state, is the time step The input feature vector is , , , are the weight matrices of the input gate, forget gate, output gate, and memory unit, respectively. , , , are the weight matrices from the previous hidden state to the input gate, forget gate, output gate, and memory unit, respectively. , , , are the bias vectors of the input gate, forget gate, output gate and memory unit respectively, is the Sigmoid activation function, is the hyperbolic tangent activation function, Represents element-wise multiplication.

[0125] In step 3, the self-attention mechanism is introduced in the LSTM layer as follows:

[0126] Calculate the attention weights between time steps: ,in is the time step For time step The attention score, is the weight vector, is the transpose of the weight vector, and are the weight matrices of the hidden states of the current time step and all time steps, respectively. is the bias vector;

[0127] Calculate the normalized attention weights: ,in is the time step For time step The attention weight of

[0128] Calculate the attention-weighted context vector: , is the time step The attention-weighted context vector of ;

[0129] Combine the attention-weighted context vector with the hidden state of the LSTM, , is the time step The final output of

[0130] and are the context vector and the weight matrix of the hidden state respectively; use the feature data set extracted and selected in step 2 to train a multi-level model combining a convolutional neural network CNN and a long short-term memory network LSTM, and introduce a self-attention mechanism; optimize the model parameters so that the model can accurately capture the nonlinear and time-varying characteristics of power demand; use the validation data set to evaluate the model performance, ensure the generalization ability of the model, and use the cross-validation method to prevent overfitting;

[0131] In step 3, a multi-level model combining CNN and LSTM is constructed, and a self-attention mechanism is introduced in the LSTM layer to dynamically adjust the model's attention to features at different time steps; this can more accurately capture the nonlinear and time-varying characteristics of electricity demand.

[0132] Step 4. Introduce adaptive reinforcement learning to obtain real-time power demand and environmental data from the smart grid and update the dataset in real time;

[0133] An adaptive reinforcement learning model is built based on the deep Q network DQN to dynamically adjust the demand response strategy; the model self-adjusts and optimizes according to real-time data to improve the flexibility and effectiveness of demand response;

[0134] Use the prediction results of the multi-level deep learning model as the input of the reinforcement learning model to optimize the strategy and ensure the accuracy of demand response decisions;

[0135] In step 4, an adaptive reinforcement learning model based on the deep Q network DQN is constructed:

[0136] Define states, actions, and rewards:

[0137] state : Current power demand and environmental data, including power consumption, temperature, and humidity;

[0138] action : Demand response strategies, including increasing and decreasing load;

[0139] award : Grid stability and economic benefits after demand response, including load balancing and cost savings;

[0140] Q value update formula: ,in is the state action value function, which means that in the state Take Action Expected rewards; is the learning rate, It is a timely reward, reflecting the effectiveness of the current demand response strategy. is the discount factor, which indicates the influence of future rewards on current decisions. The next state The maximum Q value of

[0141] Deep Q network structure:

[0142] Input layer: input current state ;

[0143] Hidden layer: multiple fully connected layers, using the activation function ReLU to improve the nonlinear representation ability of the model;

[0144] Output layer: outputs the Q value of each action;

[0145] The training process is:

[0146] Experience Replay: Storing State Transition Samples , randomly select small batches of samples for training;

[0147] Target network: Use the target network to improve training stability and update the target network parameters every fixed number of steps;

[0148] Obtain real-time power demand and environmental data from the smart grid and build an adaptive reinforcement learning model based on the deep Q network DQN; the model self-adjusts and optimizes according to real-time data, dynamically adjusts demand response strategies, and improves the stability and economic benefits of the power grid;

[0149] Step 5. Conduct demand response potential assessment, using the trained multi-level deep learning model and adaptive reinforcement learning model to predict electricity demand and identify potential demand response opportunities;

[0150] Combine historical and real-time data to analyze the impact of various factors on electricity demand and assess the demand response potential under different circumstances;

[0151] Select key evaluation indicators, including prediction accuracy, response time, and user engagement, to comprehensively evaluate the potential of demand response;

[0152] Identify the optimal demand response solution by comparing the effects of different strategies;

[0153] The specific implementation methods for predicting electricity demand and identifying potential demand response opportunities are:

[0154] Use the trained multi-layer deep learning model to predict future electricity demand;

[0155] The forecast results include short-term (hourly level), medium-term (daily level) and long-term (weekly or monthly level) electricity demand;

[0156] Identify peak load periods and off-peak load periods by analyzing forecasted electricity demand data;

[0157] During peak load periods, identify time windows where demand response can be implemented and assess the potential of demand response;

[0158] The specific implementation method for analyzing the impact of various factors on electricity demand is:

[0159] Use historical electricity demand data to analyze the impact of various factors (weather, economic activities, user behavior) on electricity demand;

[0160] Use statistical methods to quantify the impact of various factors on electricity demand;

[0161] Utilize real-time data to analyze the impact of current environment and demand changes on electricity demand;

[0162] Compare real-time data with historical data, identify abnormal changes, and adjust demand response strategies;

[0163] Combine historical and real-time data to conduct comprehensive analysis and evaluate the demand response potential under different scenarios.

[0164] Among the existing solutions, the support vector machine (SVM) machine learning model is usually used to process nonlinear relationships through kernel functions, but its performance is limited when dealing with time-varying characteristics; while some use the random forest integration method based on multiple decision trees, which can handle certain nonlinearities, but its ability to process time series data is insufficient;

[0165] Based on the existing technology, the present invention further improves the ability to capture and predict power demand fluctuations with nonlinear and time-varying characteristics by combining a multi-level deep learning model: traditional statistical analysis methods are difficult to effectively capture the nonlinear and time-varying characteristics of power demand. Although linear regression, time series analysis and other methods have certain effects when dealing with single factors or short-term predictions, it is difficult to extract key features when faced with multi-source data and complex influencing factors. The improved near-zero energy building power grid demand response potential assessment method provided by the present application extracts local features from multi-source data through a convolutional neural network CNN, combines a variational autoencoder VAE to perform feature dimensionality reduction and L1 regularization to select features, making the feature data set more representative and highly interpretable.

[0166] The present invention uses a convolutional neural network (CNN) to extract local features from multi-source data, extracting important information including temperature change patterns and load fluctuation characteristics; extracting key time series features such as daily peak and trough electricity consumption through time series slicing, and analyzing the periodicity and trend characteristics of electricity consumption data; using a variational autoencoder (VAE) to perform feature dimensionality reduction, remove redundant information, retain important features, and select the most representative features through the L1 regularization method, so that the feature data set used for demand response evaluation is highly representative and highly interpretable;

[0167] In the multi-level deep learning model construction stage, a multi-level model combining CNN and long short-term memory network LSTM is constructed. CNN is used to extract local features, and LSTM is used to capture long-term dependencies. The self-attention mechanism is introduced in the LSTM layer to dynamically adjust the model's attention to features at different time steps, so that the model can more accurately capture the nonlinear and time-varying characteristics of power demand. The feature data set extracted and selected in step 2 is used for training, and the cross-validation method is used to evaluate the model performance to prevent overfitting.

[0168] Adaptive reinforcement learning is introduced to obtain real-time power demand and environmental data through smart grids, and an adaptive reinforcement learning model based on the deep Q network DQN is constructed to dynamically adjust the demand response strategy. The model adjusts and optimizes itself according to real-time data to improve the flexibility and effectiveness of demand response. The DQN model optimizes the strategy by defining states, actions, and rewards, and improves the training stability and effectiveness through experience replay and target network update mechanisms.

[0169] During the demand response potential assessment phase, the trained multi-level deep learning model and adaptive reinforcement learning model are used to predict future electricity demand and identify potential demand response opportunities. The impact of various factors on electricity demand is analyzed by combining historical data and real-time data, and the demand response potential under different circumstances is assessed. The optimal demand response solution is identified by comparing the effects of different strategies.

[0170] It not only improves the accuracy of electricity demand forecasting, but also optimizes the execution effect of demand response strategies, effectively promoting the stable operation of smart grids and efficient use of resources.

[0171] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation manner. Any other implementation manners derived by those skilled in the art based on the technical solution of the present invention also fall within the scope of protection of the present invention.

Claims

1. A method for evaluating the demand response potential of a near-zero energy building power grid, characterized in that: The following steps are involved: Step 1. Grid demand data collection: collect power load data, weather data, economic activity data, and user behavior data; process missing values, outliers, and noise data, and standardize and normalize the collected data. Step 2: Extract the features of grid demand data. Use the convolutional neural network (CNN) to extract local features from multi-source data, extract temperature change patterns from weather data, extract load fluctuation features from power consumption data, slice the time series data, and extract key time series features, including the peak and valley time periods of daily power consumption. Use variational autoencoder VAE to reduce feature dimensionality, remove redundant information and retain important features. In step 2, a CNN model is constructed, and then CNN model training and feature extraction are performed. The CNN model architecture includes: Input layer: input preprocessed multi-source data, Convolutional layer: deploy multiple convolutional layers and use convolution kernels of different sizes to extract local features. Activation function: A non-linear activation function is applied after the convolutional layer. Pooling layer: Add pooling layer as needed to reduce feature dimensions. Fully connected layer: Map the pooled features to the fully connected layer to obtain the feature vector, and then train the CNN model to extract the temperature change pattern from the weather data and the load fluctuation characteristics from the electricity consumption data. In step 2, the time series data is sliced ​​to extract key time series features: The power load data is sliced ​​at fixed time intervals to form time series segments, and key time series features are extracted, including daily peak power consumption, that is, identifying the time period with the highest daily power consumption, and daily trough power consumption, that is, identifying the time period with the lowest daily power consumption. The periodicity and trend characteristics of power consumption data are analyzed, including the difference in power consumption patterns on weekdays and weekends. The variational autoencoder VAE is used for feature dimensionality reduction. The VAE model architecture includes: Input layer: Input high-dimensional features extracted from multi-source data. Encoder: maps high-dimensional input features to a low-dimensional latent space, Decoder: reconstructs the original features from the low-dimensional latent space, Loss function: including reconstruction loss and KL divergence, Train the VAE model to map high-dimensional features to low-dimensional latent space and remove redundant information. In step 2, the L1 regularization method is used to select the most representative features: Select linear regression, logistic regression and Lasso regression to build a model with L1 regularization, add L1 regularization terms to the loss function, input the reduced-dimensional features into the model, and train and evaluate the model with L1 regularization. Through L1 regularization, remove features with weights close to zero. Step 3. Multi-level deep learning model construction: construct a multi-level model combining CNN and long short-term memory network LSTM. CNN is used to extract local features, LSTM is used to capture long-term dependencies, and the self-attention mechanism is introduced in the LSTM layer to dynamically adjust the model's attention to features at different time steps. The multi-level deep learning model is trained using the feature dataset extracted and selected in step 2, and the model performance is evaluated using the cross-validation method. In step 3, the multi-level model architecture combining CNN and LSTM is as follows: Input layer: input preprocessed multi-source data, where ,in is the number of samples, is the number of time steps, is the number of features; Convolutional layer CNN: Use multiple convolutional layers to extract local features: in, It is The output feature map of the layer, It is Tier convolution kernels, It is The bias of the layer, is the ReLU activation function, which represents the convolution operation; Pooling layer: Apply pooling operation to reduce feature dimension: , It is The pooled output feature map of the layer, is the pooling function, which expands the pooled feature map into a time step feature sequence ,in It is the output feature dimension of the pooling layer; Long Short-Term Memory Network LSTM layer: , , , , ,in, are the activation vectors of the input gate, forget gate, and output gate, respectively. is the memory unit state; is the hidden state, is the time step The input feature vector is They are the weight matrices of the input gate, forget gate, output gate, and memory unit. are the weight matrices from the previous hidden state to the input gate, forget gate, output gate, and memory unit, respectively. are the bias vectors of the input gate, forget gate, output gate and memory unit respectively, is the Sigmoid activation function, is the hyperbolic tangent activation function, represents element-wise multiplication, Introduce the self-attention mechanism in the LSTM layer to calculate the attention weights between time steps: ,in, is the time step For time step The attention score, is the weight vector, is the transpose of the weight vector, and are the weight matrices of the hidden states of the current time step and all time steps, respectively. is the bias vector; Calculate the normalized attention weights: ,in, is the time step For time step The attention weight, Calculate the attention-weighted context vector: in, is the time step The attention-weighted context vector of Combine the attention-weighted context vector with the hidden state of the LSTM: ,in, is the time step The final output is and are the weight matrices of the context vector and hidden state, respectively. The feature dataset extracted and selected in step 2 is used to train a multi-level model combining CNN and LSTM, and a self-attention mechanism is introduced. Step 4. Introduce adaptive reinforcement learning, obtain real-time power demand and environmental data from the smart grid, update the data set in real time, build an adaptive reinforcement learning model based on the deep Q network DQN, dynamically adjust the demand response strategy, and the model self-adjusts and optimizes according to real-time data. The prediction results of the multi-level deep learning model are used as the input of the reinforcement learning model. In step 4, an adaptive reinforcement learning model based on the deep Q network DQN is constructed: Define states, actions, and rewards: Status: current power demand and environmental data, including power consumption, temperature, humidity, Actions: Demand response strategies, including increasing and decreasing load, Rewards: Grid stability and economic benefits of demand response, including load balancing and cost savings, Q value update formula: ,in, is the state action value function, which means that in the state Take Action Expected rewards; is the learning rate, is an immediate reward that reflects the effectiveness of the current demand response strategy. is the discount factor, which indicates the influence of future rewards on current decisions. The next state The maximum Q value of Deep Q network structure: Input layer: input the current state, Hidden layer: multiple fully connected layers, using the activation function ReLU to improve the nonlinear representation ability of the model. Output layer: outputs the Q value of each action, The training process includes experience replay and target network: Experience replay: store state transition samples and randomly extract small batches of samples for training. Target network: Use the target network to improve training stability and update the target network parameters every fixed number of steps. Step 5. Demand response potential assessment: Use the trained multi-level deep learning model and adaptive reinforcement learning model to predict electricity demand and identify potential demand response opportunities. Combine historical data and real-time data to analyze the impact of various factors on electricity demand and assess the demand response potential under different circumstances. Select key evaluation indicators, including prediction accuracy, response time, and user participation, to comprehensively assess the potential of demand response and identify the optimal demand response solution by comparing the effects of different strategies. The steps of predicting electricity demand and identifying potential demand response opportunities include: using a trained multi-level deep learning model to predict future electricity demand. The prediction results include short-term, medium-term and long-term electricity demand. Analyze the predicted electricity demand data, identify peak load periods and valley load periods, identify time windows for implementing demand response during peak load periods, and evaluate the potential for demand response. The steps to analyze the impact of various factors on electricity demand include: Utilize historical electricity demand data to analyze the impact of various factors on electricity demand. Use statistical methods to quantify the degree of impact of various factors on electricity demand. Utilize real-time data to analyze the immediate impact of the current environment and demand changes on electricity demand. Compare real-time data with historical data to identify abnormal changes and adjust demand response strategies. Combine historical data with real-time data to conduct a comprehensive analysis and evaluate the demand response potential under different scenarios.

Images