Building energy consumption management system and method for analyzing power grid load

Abstract

The invention relates to the technical field of building energy consumption management, in particular to a building energy consumption management system and method for analyzing the load of a power grid, and the system comprises an acquisition processing unit, a load analysis unit and an energy management unit. A load prediction module of the load analysis unit predicts a load by using a long short-term memory network and a graph convolutional neural network, a decomposition traceability module decomposes a traceability load by using an attention mechanism and a hidden Markov model, and an energy management unit regulates and controls adjustable equipment in peak and valley periods according to load prediction. According to the method and the system, the abnormal energy consumption equipment is pre-warned according to the load decomposition and traceability result, multiple technical means are integrated, the traceability load can be accurately predicted and decomposed, the energy management is optimized, the complex building environment is adapted, and the problems that the equipment relationship is difficult to grasp, the load change reason is difficult to track and the energy consumption is difficult to accurately manage in a traditional mode are effectively solved.

Description

Technical Field

[0001] Building Energy Consumption Management The present invention relates to the technical field, and in particular to a building energy consumption management system and method for analyzing power grid load. Background Art

[0002] Building energy management is an important technology. In the operation of modern buildings, the number of electrical equipment continues to increase and the types are diverse, covering various facilities such as lighting systems, air-conditioning systems, electrical equipment, office appliances, etc. Their operating modes and energy consumption characteristics vary greatly. At the same time, the meteorological conditions of the building environment also have an important effect on the energy consumption of equipment.

[0003] This complex interaction causes the grid load to show frequent fluctuations, and its changing patterns are difficult to grasp, which in turn brings huge challenges to load forecasting. Traditional building energy consumption management methods often rely on simple statistical analysis and empirical judgment, and can only monitor equipment energy consumption in isolation. It is unable to deeply understand and grasp the comprehensive impact of the relationship between equipment on the overall load. In addition, when the grid load changes, the traditional method cannot trace back to the specific equipment and reasons that cause the load change in time due to the lack of effective analysis methods. It cannot accurately identify whether the load fluctuation is caused by a failure of a certain equipment, an adjustment of the operating status, or a change in environmental factors. This leads to serious limitations in optimizing energy consumption, making it difficult to formulate targeted energy-saving measures, and thus it is impossible to achieve accurate management of building energy consumption. In order to solve this problem, we provide a building energy consumption management system and method for analyzing grid load. Summary of the invention

[0004] The object of the present invention is to provide a building energy consumption management system and method for analyzing power grid load, so as to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above-mentioned purpose, a building energy consumption management system for analyzing power grid load is provided, including a collection and processing unit, a load analysis unit and an energy management unit;

[0006] The collection and processing unit is used to collect the operation data of the electrical equipment in the building and the meteorological data of the environment in which the building is located, and pre-process the operation data and the environmental data, align the equipment operation data and the meteorological data according to the time series, and then construct the connection relationship data between the equipment into graph structure data, in which the equipment is used as the node of the graph and the connection relationship is used as the edge of the graph;

[0007] The load analysis unit includes a load forecasting module and a decomposition and tracing module. The load forecasting module takes the preprocessed time series data as input, captures the dependency of the data in the time dimension through the long short-term memory network, and then extracts features from the graph structure data through the graph convolutional neural network to learn the impact of the relationship between devices on the load and obtain the predicted power grid load. The decomposition and tracing module uses the attention mechanism to assign weights to the equipment operation data, and combines the hidden Markov model to model the load mode of the equipment under different states, which is used to decompose the current power grid load on each device and trace the load change.

[0008] The energy management unit issues instructions to adjust operating parameters to adjustable equipment in the building during peak and valley periods of grid load according to the predicted grid load output by the load prediction module, and issues fault warnings to abnormal energy consumption equipment based on load decomposition and tracing results.

[0009] As a further improvement of the technical solution, the acquisition processing unit makes adjustments according to the differences in the connection relationships between devices when constructing the graph structure data, as follows:

[0010] Identify and classify each type of equipment relationship. For electrical connection relationships, they are divided into strong current connection and weak current connection. For communication connection relationships, they are divided into wired communication and wireless communication.

[0011] Then, a weight value is assigned to each connection relationship based on the data transmission volume of the connection relationship. When constructing graph structure data, the weight value is used as the attribute of the edge in the graph. The load analysis unit evaluates the impact of different connection relationships on equipment operation and energy consumption based on the weight attribute of the edge, and the acquisition and processing unit updates the weight value of the connection relationship in real time according to changes in the equipment operation status.

[0012] As a further improvement of the technical solution, the long short-term memory network in the load forecasting module inputs the time series data time step by time step, as follows:

[0013] The long short-term memory network is composed of a forget gate, an input gate, a cell state and an output gate;

[0014] The forget gate integrates information based on the input data of the current time step and the hidden state of the previous moment to obtain a vector, and inputs the vector into a sigmoid activation function. The value output by the sigmoid activation function determines whether the element in the cell state of the previous moment is retained or forgotten;

[0015] The input gate is divided into two parts. The first part performs the same operation as the forget gate. The second part performs a linear transformation on the current input data and the hidden state at the previous moment to generate a candidate value vector. The two parts are multiplied to obtain a new cell state.

[0016] Multiply the vector of the forget gate by the cell state at the previous moment element by element, forget the information in the cell state at the previous moment, and then add the new cell state generated by the input gate to the retained element to perform a second update of the cell state;

[0017] The output gate is calculated based on the cell state after the second update and the hidden state at the previous moment to obtain the hidden state output of the current time step. The hidden state output is used to capture the dependency of time series data in the time dimension, and the initial hidden state of the long short-term memory network is adaptively initialized according to the characteristics of the time series data first received by the system.

[0018] As a further improvement of the technical solution, the operation of the graph convolutional neural network in the load forecasting module when extracting features from graph structure data is as follows:

[0019] An adjacency matrix is ​​constructed based on the connection relationship between devices in the acquisition and processing unit to represent the connection between nodes in the graph. At the same time, the device power, operating time and device type of each node are used as the feature matrix;

[0020] Initialize the dimension of the convolution kernel, the convolution kernel belongs to the graph convolutional neural network itself, the convolution kernel dimension matches the dimension of the node feature matrix, and the initialization weight of the convolution kernel is randomly assigned based on the distribution of node features;

[0021] The convolution kernel performs a sliding convolution operation on the graph structure data. For each node, the convolution kernel determines its neighbor nodes based on the adjacency matrix, and then multiplies its own weight with the eigenvalues ​​of the central node and its neighbor nodes and sums them up to aggregate local features, obtain the impact of the relationship between devices on the load, and dynamically adjust the calculation weight of the convolution kernel on the node according to the contribution of the node during the convolution process.

[0022] As a further improvement of the technical solution, the graph convolutional neural network in the load forecasting module extracts the relationship features between devices for load forecasting during the convolution operation process, as follows:

[0023] For the first layer of the graph convolutional neural network, the convolution kernel function is determined according to the number of nodes, connection density and initial features of the nodes in the initial graph structure. If the number of nodes is N, the connection density is D, the feature dimension is F, the convolution kernel size is set to k1×k1, the step size is set to s1, the padding method is p1 and the number of convolution kernels is set to c1;

[0024] For each node, the convolution kernel determines its neighboring nodes based on the adjacency matrix, multiplies the convolution kernel weight with the eigenvalues ​​of the central node and its neighboring nodes, and sums them up, completing local feature aggregation and extracting the primary features of the relationship between the node and its neighboring nodes. After the first layer of convolution, the nodes are clustered according to the extracted features.

[0025] In the second layer, if the first layer output feature dimension is F 1 , the convolution kernel size is k2×k2, (k2>k1), the step size is s2, the padding method is p2, and the convolution operation is performed based on the adjusted parameters to fuse the local relations extracted by the first layer, update the primary node features, obtain the intermediate node features, and introduce the skip connection in the second layer of convolution to directly transfer the feature information of the first half of the first layer to the second layer;

[0026] For the third layer, if the second layer output feature dimension is F 2 , the convolution kernel size is set to k3×k3, (k3>k2), the step size is set to s3, the padding method is set to p3, the convolution kernel weight is updated according to the output features of the second layer using the adaptive learning rate adjustment algorithm, and the number of convolution kernels is increased to c3;

[0027] The adjusted parameters are used for convolution to obtain the load relationship characteristics between the devices in the graph, and load prediction is performed based on the load relationship characteristics between the devices in the graph.

[0028] As a further improvement of the technical solution, the operation of the attention mechanism in the analysis and traceability module when weighting the device operation data is assigned is as follows:

[0029] Perform linear transformation on the equipment operation data, set the equipment operation data matrix X, and multiply the equipment operation data matrix by the preset weight matrix W q , W k and W v , generate query vector, key vector and value vector respectively;

[0030] The similarity between the query vector and the key vector is calculated to form a similarity matrix, and the similarity matrix is ​​normalized to obtain weights. The weights are used to perform weighted summation on the value vectors, and the weighted summation is used to assign weights to the device operation data.

[0031] As a further improvement of the technical solution, the operation of the hidden Markov model in the decomposition and tracing module when modeling the composite mode under different states of the device is as follows:

[0032] Extract the power, current, voltage and operating frequency of the equipment from the acquisition unit, and extract the mean value and the statistical characteristics of the correlation between the parameters;

[0033] Obtain the equipment's past operating status records and load data from the historical database, compare and integrate the current data features with them, and identify the operating mode and status change trend;

[0034] Based on the above analysis, the equipment operation conditions with the same characteristics and historical performance are classified into categories, and each category is an equipment state, including normal operation, different load levels, and fault operation states;

[0035] For each equipment state, the occurrence frequency of different load value ranges in this state is counted, and a normal distribution is constructed to describe the load pattern. The parameters of the probability distribution function are determined by maximum likelihood estimation.

[0036] When new equipment operation data is transmitted, the Bayesian theorem method is used to calculate the probability of occurrence in each state by comparing the new data with the characteristics of each state and the probability distribution of the load pattern, so as to determine the current state of the equipment, perform load decomposition and trace the source, and introduce prior knowledge to correct the probability calculation.

[0037] As a further improvement of the technical solution, during the operation of the hidden Markov model in the decomposition and tracing module, the probability distribution of the load mode in each state is updated in real time according to the newly acquired equipment operation data, as follows:

[0038] When new equipment operation data is obtained, for each determined equipment state in the model, the load value under that state is re-counted, and the load pattern probability distribution under each state is updated in real time according to the value;

[0039] The device was in state S at the previous moment. a , in state S at the current moment b The probability is the state transition probability. According to Bayes' theorem, after new data is acquired, these probability values ​​are recalculated, and the state transition probability is adjusted through the calculation.

[0040] A second object of the present invention is to provide a method for implementing a building energy consumption management system for analyzing power grid loads as described in any one of the above, comprising the following steps:

[0041] S1. Extract the power, current and voltage of the electrical equipment in the building and the temperature and humidity data of the building environment, align the equipment operation data and meteorological data in time series, and construct the connection relationship data between devices into graph structure data;

[0042] S2. First, input the preprocessed time series data into the long short-term memory network step by time, and capture the time dependency through the coordinated operation of the forget gate, input gate and output gate. Then, input the data processed by the long short-term memory network and the graph structure data into the graph convolutional neural network. The graph convolutional neural network constructs the adjacency matrix and feature matrix according to the device connection relationship. After adjusting the multi-layer convolution kernel parameters, the local features of the nodes are gradually extracted to the load relationship features between devices to perform load forecasting.

[0043] S3. Use the attention mechanism to linearly transform the device operation data matrix and generate a vector with the preset weight matrix, calculate the similarity, normalize the weight to obtain the weight to assign the data weight, and combine the hidden Markov model to extract the device power characteristics and integrate historical data to determine the state and load mode probability distribution. When new data is passed in, use the Bayesian theorem to calculate the probability to determine the state to achieve decomposition and traceability. During operation, the load mode probability distribution and state transition probability of the state are also updated in real time;

[0044] S4. Based on the predicted grid load, the time period is determined by comparing it with the peak and valley thresholds set based on the historical energy consumption data of the building and the power supply capacity of the grid. During the peak and valley periods, instructions for adjusting operating parameters are issued to the adjustable equipment. At the same time, based on the load decomposition and tracing results, the probability distribution of the equipment energy consumption pattern and the normal load pattern are compared, and fault warnings are issued for abnormal energy consumption equipment.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The load prediction module and the decomposition and tracing module in the present invention work together and promote each other. The prediction results of the load prediction module provide the decomposition and tracing module with an expected framework of future load changes, enabling it to plan analysis strategies in advance. The relationship characteristics between devices extracted by the decomposition and tracing module during the analysis process are fed back to the load prediction module to help it optimize the learning of the relationship between devices and further improve the accuracy of load prediction. This collaborative working mechanism greatly improves the accuracy of load decomposition and tracing. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is an overall block diagram of the present invention;

[0048] Figure 2 It is the overall flow chart of the present invention.

[0049] The meaning of each number in the figure is:

[0050] 1. Collection and processing unit; 2. Load analysis unit; 21. Load forecasting module; 22. Decomposition and tracing module; 3. Energy management unit; DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] During building operation, there are many electrical equipment and their operation is complex. Their energy consumption is affected by many conditions such as the equipment's own characteristics, operating status, and environmental factors, resulting in frequent and unpredictable fluctuations in the grid load. Traditional management methods make it difficult to fully grasp the impact of the relationship between equipment on the load, and are unable to track the causes of load changes in a timely manner, which is not conducive to optimizing energy consumption.

[0053] The present invention provides a building energy consumption management system for analyzing power grid load. Figure 1 As shown, it includes a collection and processing unit 1, a load analysis unit 2 and an energy management unit 3;

[0054] The collection and processing unit 1 collects the operating data of the electrical equipment in the building and the meteorological data of the environment in which the building is located, and preprocesses the operating data and environmental data, and inputs the preprocessed time series data as input to the long short-term memory network in the load forecasting module 21, and captures the dependency of the data in the time dimension through the long short-term memory network, as follows:

[0055] The acquisition and processing unit 1 uses various sensors and data acquisition equipment to collect the operating data of the electrical equipment in the building, including but not limited to the power, current and voltage data of the equipment and the temperature and humidity of the building environment. At the same time, it collects the connection relationship data between the devices, including the electrical connection type, communication connection method and the data transmission volume of the connection.

[0056] For power data, if the power value at a certain moment exceeds the normal operating range, it is judged as an abnormal value and corrected by interpolation method. The interpolation method is an existing technology and the specific calculation formula is not repeated here. Data of different types and magnitudes are normalized to make the data comparable.

[0057] The preprocessed data is input into the long short-term memory network. The initial hidden state of the long short-term memory network is adaptively initialized according to the characteristics of the time series data first received by the system. For the preprocessed time series data, it is input into the long short-term memory network time step by time. At each time step t, the forget gate is based on the input data x of the current time step. t and the hidden state h at the previous moment t-1 To integrate information, please refer to the following:

[0058] Concatenate them into a vector [x t ;h t-1 ], and then multiply it by a weight matrix W f , the dimension is (d x +d h )×d h , where d x is the dimension of the input data, d h is the dimension of the hidden state, plus a bias vector b f , dimension d h , we get a vector f t =W f ×[x t ;h t-1 ]+b f ; f t Input into a sigmoid activation function σ to get the output f of the forget gate t =σ(f t ), whose element value is between 0 and 1, determines the cell state c at the previous moment t-1 The degree to which each element in China is retained or forgotten, where the calculation formula of the forget gate is expressed as f t =σ(W f ·[x t ;h t-1 ]+b f ).

[0059] The input gate is divided into two parts. The first part is the same as the forget gate operation, which converts [x t ;h t-1 ] multiplied by the weight matrix W i , the dimension is (d x +d h )×d h , plus the bias vector b i , dimension d h , and then through the sigmoid activation function, the output i of the first part of the input gate is obtained t =σ(W i ·[x t ;h t-1 ]+b i ).

[0060] The second part is the current input data x t and the hidden state h at the previous moment t-1 After linear transformation, the candidate value vector is generated, and [x t ;h t-1 ] multiplied by the weight matrix W c , the dimension is (d x +d h )×d c, where d c is the dimension of the cell state, plus the bias vector b c , and then through the tanh activation function, we get the candidate value vector

[0061] The output gate is based on the cell state c after the second update t and the previous hidden state h t-1 To calculate, first [c t ;h t-1 ] multiplied by the weight matrix W 0 , the dimension is (d c +d h )×d h , plus the bias vector b 0 , dimension d h , and then the control signal of the output gate is obtained through the sigmoid activation function

[0062] Through the above time-step operations, the long short-term memory network can capture the dependencies of time series data in the time dimension. The hidden state h t It is continuously updated at each time step, and contains the comprehensive features of historical information and current input information, providing an effective time series feature representation for subsequent load forecasting in combination with graph convolutional neural networks. At the same time, throughout the process, the parameters of the long short-term memory network can be learned and adjusted according to the training data through the back-propagation algorithm to improve the accuracy of load forecasting.

[0063] The equipment operation data and meteorological data are aligned in time series, and then the connection relationship data between devices is constructed into graph structure data, where the devices are nodes of the graph and the connection relationships are edges of the graph. The graph structure data is extracted through a graph convolutional neural network to learn the impact of the relationship between devices on the load and obtain the predicted power grid load. The details are as follows:

[0064] The adjacency matrix A is constructed based on the inter-device connection relationship data acquired by the acquisition processing unit 1. If there is a connection relationship between device i and device j, then A ij =ω ij ,ω ij is the weight value assigned according to the connection relationship attribute. If there is no connection relationship, then A ij =0; at the same time, construct the feature matrix X of each node, and combine the multi-dimensional information of each node such as device power P, running time T and device type E into a feature vector.

[0065] Initialize the dimension of the convolution kernel to match the dimension of the node feature matrix. Suppose the dimension of the node feature matrix X is n×d, then the dimension of the convolution kernel W is k×k×d×c, where k is the convolution kernel size and c is the number of convolution kernels. The initialization weight of the convolution kernel is assigned randomly or with a specific strategy based on the distribution of node features.

[0066] For the first layer of the graph convolutional neural network, the convolution kernel function is determined according to the number of nodes n, the connection density D and the initial feature dimension d of the node in the initial graph structure. For each node i, the convolution kernel determines its neighborhood node N(i) according to the adjacency matrix A, and then multiplies the convolution kernel weight W with the eigenvalues ​​of the central node i and its neighborhood node j∈N(i) and sums them to perform local feature aggregation. The specific calculation formula is: in represents the weight matrix of the lth convolution kernel, x j is the feature vector of node j, and the new feature vector y of node i is obtained i , thereby extracting the primary features of the relationship between a node and its neighboring nodes.

[0067] After the first layer of convolution, the nodes are clustered according to the extracted primary features and divided into K groups according to feature similarity. K can be determined based on factors such as the functional area or energy consumption pattern of the equipment in the building, so as to analyze the relationship between equipment at different levels in the future.

[0068] If the first layer output feature dimension is d 1 , where d 1 =C 1 , that is, the number of convolution kernels, then the size of the second layer convolution kernel is k2, the step size is s2, and the filling method P2 is determined according to k2 and s2. At the same time, the convolution operation is performed based on the adjusted parameters, the local relations extracted by the first layer are integrated, the primary node features are updated, and the intermediate node features are obtained.

[0069] Introduce skip connections in the second convolution layer to directly transfer part of the feature information of the first layer to the second layer. The specific approach is as follows:

[0070] The feature vector y output by the first layer i The new feature vector z calculated by the second layer convolution i Concatenate and get the updated feature vector u i =[y i ;z i ]; This preserves the local information extracted by the first layer while enhancing the second layer’s ability to learn relationships between devices in a wider area.

[0071] If the second layer output feature dimension is d 2 , d 2According to the number of convolution kernels in the second layer and the dimension after feature fusion, the convolution kernel size is k3, the step size is s3, P3 is determined according to s3, and the convolution kernel weight is updated according to the output features of the second layer using an adaptive learning rate adjustment algorithm. The learning rate is dynamically adjusted according to the gradient information of the features, so that the convolution kernel can better adapt to the data features. At the same time, the number of convolution kernels is increased to c3.

[0072] The adjusted parameters are used for convolution to obtain the load relationship characteristics between the devices in the graph. Finally, load prediction is performed based on the load relationship characteristics between the devices in the graph. The fully connected layer is used to map the features output by the convolution layer to the predicted load value, as follows:

[0073] Let the weight matrix of the fully connected layer be W f , the dimension is m×1, m is the feature dimension of the last convolution output, and the bias vector is b f , then the predicted load value Where v is the feature vector output by the last convolution layer.

[0074] After the third layer of convolution, the extracted features are visualized for analysis and debugging. The principal component analysis dimensionality reduction method can be used to project the high-dimensional feature vector into two-dimensional or three-dimensional space, draw feature graphs, observe the distribution and clustering of features between devices, and determine whether they meet the expected device relationship pattern. It also helps to discover abnormal device features and relationships and further optimize the model.

[0075] The load forecasting module 21 predicts the peak load period and approximate load amount for each day in the next week. The decomposition and tracing module 22 uses this information to analyze in advance the possible operating status and load contribution mode of each device under these expected load conditions, so as to better perform equipment status assessment and anomaly detection.

[0076] Among them, in the process of analyzing equipment operation data and building equipment status model, the decomposition and tracing module 22 can extract the relationship characteristics between equipments and feed back to the load forecasting module 21. The load forecasting module 21 uses this information to optimize its learning of the impact of equipment relationships on load, and further improve the accuracy of load forecasting.

[0077] The decomposition and tracing module 22 uses the attention mechanism to assign weights to the equipment operation data, and combines the hidden Markov model to model the load patterns of the equipment under different states, which is used to decompose the current power grid load on each device and trace the load changes, as follows:

[0078] Set the device operation data matrix JZ and multiply it by the preset weight matrix W q , with dimensions of d×d q , where d is the characteristic dimension of the equipment operation data, d qis the expected dimension of the query vector, W k and W v Generate query vector Q, key vector K and value vector V respectively, the calculation formula is Q = JZ × W q , K=JZ×W k , V = XW v .

[0079] Calculate the similarity between the query vector and the key vector, and use the dot product to calculate the similarity matrix S, that is, S ij =Q i ×K j ; where Q i is the i-th row of the query vector Q, K j is the j-th row of the key vector K.

[0080] The similarity matrix S is normalized to obtain the weight ωA, and the softmax function is used for normalization, then These weights are used to perform weighted summation on the value vector V to obtain the attention-weighted device operation data representation Through the attention mechanism, weights can be assigned according to the importance of different features of device operation data, highlighting key information and providing more targeted data for subsequent hidden Markov model modeling.

[0081] According to the extracted equipment operation data characteristics and historical data comparison and integration results, the equipment operation states with the same characteristics and historical performance are classified. For each equipment state, the frequency of occurrence of different load value ranges in this state is counted, and then a probability distribution function is constructed based on this. If the data conforms to the normal distribution characteristics, let the load be L fh , then the normal distribution probability density function Where μ and σ are the mean and standard deviation of the normal distribution respectively.

[0082] When new equipment operation data is passed in, the Bayesian theorem method is used to calculate its probability of occurrence in each state, as follows:

[0083] The device has m states, S 1 , S 2 , ..., S m , the new data is x, the prior probability P(S i ), conditional probability P(x|S i ) is calculated based on the probability distribution of load patterns under each state currently constructed, then the posterior probability If the probability of failure of certain equipment is higher in a specific season or time period, the prior probability of the corresponding fault state should be appropriately increased. At the same time, the prior probability and conditional probability should be adjusted in combination with the load forecast results. If it is predicted that the load of a certain equipment will fluctuate greatly in the future, then when calculating the probability of it being in an abnormal state, the prior probability of the relevant abnormal state should be appropriately increased, and the calculation of the conditional probability should be adjusted according to the predicted load range.

[0084] The load pattern probability distribution in each state is updated in real time according to the newly acquired equipment operation data. When new data is obtained, for each determined equipment state in the model, the load value in that state is re-counted.

[0085] If under normal operating conditions, the number of load occurrences in a certain interval in the new data increases, the probability value of the interval will be adjusted accordingly to make the probability distribution more in line with the actual situation. The update frequency can be dynamically adjusted according to the stability of the equipment operation and the sensitivity of data changes. If the equipment operation is relatively stable and the data changes slowly, the update frequency is set to once every hour or every day. If the equipment operation status fluctuates greatly or the data changes sensitively, the update frequency can be increased to once every minute or every few minutes.

[0086] Calculate the state transition probability P(S b |S a ), that is, the device is in state S at the previous moment a , in state S at the current moment b , according to Bayesian theorem, combined with the newly acquired data, recalculation is performed. At the same time, the continuity of equipment operation and the impact of environmental factors on state transition, as well as the trend information in the load forecast results, are considered. For example, if the equipment runs continuously for a period of time in a high temperature environment, the probability of transitioning from a normal operating state to an overload state may increase. If the load forecast results show that the overall load will increase in the future, the probability of the equipment transitioning to a high load state will also be appropriately increased. In addition, the state transition probability can be smoothed according to the historical state transition data to avoid drastic changes in the state transition probability due to fluctuations in individual data, as follows:

[0087] P new (S b |S a )=αP(S b |S a )+(1-α)P old (S b |A a );

[0088] Where α is the smoothing coefficient, P new (S b |S a ) is the updated state transition probability, Pold (S b |S a ) is the state transition probability calculated last time.

[0089] According to the current state of the equipment and its load pattern probability distribution determined by the hidden Markov model, the current grid load is decomposed into each device. If a device is currently judged to be in normal operation and its load in this state follows the normal distribution N(μ, σ 2 ), the contribution ratio of the equipment under the current load is calculated according to the probability density function of the normal distribution, as follows:

[0090] Assume that the current total grid load is L total , the contribution of the device L i Through L i =L total ×P(L i ) calculation, where P(L i ) is calculated based on the normal distribution. The equipment load is L i The probability of failure is calculated and similar calculations are performed on all devices to achieve the decomposition of the current grid load on each device. During the decomposition process, the load forecast results are referred to. For devices that contribute more to the predicted future load, more attention is paid to the current decomposition, and the difference between their current state and the predicted state is analyzed to detect potential problems in a timely manner.

[0091] When the load changes, the cause of the load change is traced by analyzing the changes in the status of each device and the changes in the load pattern probability distribution. If the status of a device changes from normal operation to overload status, and the load mean of its load pattern probability distribution increases, it can be determined that the status change of the device is an important reason for the load increase. At the same time, combined with the graph structure data between devices, it is analyzed whether other devices should be affected by the status change of the device, and the propagation path and root cause of the load change are further traced.

[0092] If the overload device is electrically connected to other devices and shares the power supply line, it may cause voltage fluctuations in other devices, thereby affecting their operating status and load. In this way, the reasons for the load changes can be fully traced to provide a decision-making basis for energy management. During the tracing process, combined with the load forecast results, it is judged whether the current load changes are in line with expectations. If not, further analysis is conducted to determine which devices’ actual operating conditions deviate from the forecast and identify the factors that cause the deviation, so as to optimize the load forecast model and energy management strategy.

[0093] Based on the load decomposition and tracing results, fault warnings are issued for abnormal energy consumption equipment, which can timely discover potential problems of equipment. Early detection and treatment of equipment failures can avoid the expansion of failures, reduce equipment downtime, ensure the normal operation of various equipment in the building, and thus maintain the stability of the entire building energy management system.

[0094] The energy management unit 3 issues instructions to adjust the operating parameters of the adjustable equipment in the building during the peak and valley periods of the grid load according to the predicted grid load output by the load prediction module 21, and at the same time issues fault warnings for abnormal energy consumption equipment based on the load decomposition and tracing results, thereby realizing intelligent management of building energy consumption. The system no longer relies on manual experience to adjust equipment operation and troubleshoot, and can automatically make decisions based on the prediction and analysis results, thereby improving the accuracy and timeliness of management.

[0095] In the present invention, the collection and processing unit collects and preprocesses the electrical equipment operation and environmental meteorological data to construct graph structure data. The load prediction module of the load analysis unit uses long short-term memory network and graph convolutional neural network to predict the load. The decomposition and tracing module uses attention mechanism and hidden Markov model to decompose the traceable load. The energy management unit regulates the adjustable equipment during peak and valley periods according to the load forecast, and warns of abnormal energy consumption equipment according to the load decomposition and tracing results. By integrating multiple technical means, it can accurately predict, decompose and trace the load, optimize energy management, adapt to complex building environments, and effectively solve the problem that traditional methods are difficult to grasp the relationship between equipment, track the causes of load changes, and accurately manage energy consumption.

[0096] A second object of the present invention is to provide a method for implementing a building energy consumption management system for analyzing power grid loads including any one of the above, comprising the following steps:

[0097] S1. Extract the power, current and voltage of the electrical equipment in the building and the temperature and humidity data of the building environment, align the equipment operation data and meteorological data in time series, and construct the connection relationship data between devices into graph structure data;

[0098] S2. First, input the preprocessed time series data into the long short-term memory network step by time, and capture the time dependency through the coordinated operation of the forget gate, input gate and output gate. Then, input the data processed by the long short-term memory network and the graph structure data into the graph convolutional neural network. The graph convolutional neural network constructs the adjacency matrix and feature matrix according to the device connection relationship. After adjusting the multi-layer convolution kernel parameters, the local features of the nodes are gradually extracted to the load relationship features between devices to perform load forecasting.

[0099] S3. Use the attention mechanism to linearly transform the device operation data matrix and generate a vector with the preset weight matrix, calculate the similarity, normalize the weight to obtain the weight to assign the data weight, and combine the hidden Markov model to extract the device power characteristics and integrate historical data to determine the state and load mode probability distribution. When new data is passed in, use the Bayesian theorem to calculate the probability to determine the state to achieve decomposition and traceability. During operation, the load mode probability distribution and state transition probability of the state are also updated in real time;

[0100] S4. Based on the predicted grid load, the time period is determined by comparing it with the peak and valley thresholds set based on the historical energy consumption data of the building and the power supply capacity of the grid. During the peak and valley periods, instructions for adjusting operating parameters are issued to the adjustable equipment. At the same time, based on the load decomposition and tracing results, the probability distribution of the equipment energy consumption pattern and the normal load pattern are compared, and fault warnings are issued for abnormal energy consumption equipment.

[0101] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected, and the scope of the present invention to be protected is defined by the attached claims and their equivalents.

Claims

1. A building energy consumption management system for analyzing power grid load, characterized in that: It comprises a data collection and processing unit (1), a load analysis unit (2) and an energy management unit (3); The acquisition and processing unit (1) is used to collect operating data of electrical equipment in a building and meteorological data of the environment in which the building is located, and pre-process the operating data and environmental data, align the equipment operating data and meteorological data according to time series, and then construct the connection relationship data between equipment into graph structure data, wherein the equipment is used as a node of the graph and the connection relationship is used as an edge of the graph; The load analysis unit (2) comprises a load prediction module (21) and a decomposition and tracing module (22). The load prediction module (21) takes the preprocessed time series data as input, captures the dependency of the data in the time dimension through a long short-term memory network, and then extracts features from the graph structure data through a graph convolutional neural network to learn the impact of the relationship between devices on the load and obtain the predicted power grid load. The decomposition and tracing module (22) uses an attention mechanism to assign weights to the device operation data and combines a hidden Markov model to model the load mode of the device under different states, so as to decompose the current power grid load on each device and trace the load change. The energy management unit (3) issues instructions to adjust operating parameters to adjustable equipment in the building during peak and valley periods of grid load according to the predicted grid load output by the load prediction module (21), and issues fault warnings to abnormal energy consumption equipment based on load decomposition and tracing results.

2. A building energy consumption management system for analyzing power grid load according to claim 1, characterized in that: When constructing the graph structure data, the acquisition processing unit (1) makes adjustments according to the differences in the connection relationships between the devices, as follows: Identify and classify each type of equipment relationship. For electrical connection relationships, they are divided into strong current connection and weak current connection. For communication connection relationships, they are divided into wired communication and wireless communication. Then, a weight value is assigned to each connection relationship according to the data transmission volume of the connection relationship. When constructing graph structure data, the weight value is used as the attribute of the edge in the graph. The load analysis unit (2) evaluates the impact of different connection relationships on equipment operation and energy consumption according to the weight attribute of the edge, and the collection and processing unit (1) updates the weight value of the connection relationship in real time according to the change of the equipment operation status.

3. A building energy consumption management system for analyzing power grid load according to claim 1, characterized in that: The long short-term memory network in the load forecasting module (21) inputs the time series data step by time, as follows: The long short-term memory network is composed of a forget gate, an input gate, a cell state and an output gate; The forget gate integrates information based on the input data of the current time step and the hidden state of the previous moment to obtain a vector, and inputs the vector into a sigmoid activation function. The value output by the sigmoid activation function determines whether the element in the cell state of the previous moment is retained or forgotten; The input gate is divided into two parts. The first part performs the same operation as the forget gate. The second part performs a linear transformation on the current input data and the hidden state at the previous moment to generate a candidate value vector. The two parts are multiplied to obtain a new cell state. Multiply the vector of the forget gate by the cell state at the previous moment element by element, forget the information in the cell state at the previous moment, and then add the new cell state generated by the input gate to the retained element to perform a second update of the cell state; The output gate is calculated based on the cell state after the second update and the hidden state at the previous moment to obtain the hidden state output of the current time step. The hidden state output is used to capture the dependency of time series data in the time dimension, and the initial hidden state of the long short-term memory network is adaptively initialized according to the characteristics of the time series data first received by the system.

4. A building energy consumption management system for analyzing power grid load according to claim 1, characterized in that: The operation of the graph convolutional neural network in the load prediction module (21) when extracting features from graph structure data is specifically as follows: An adjacency matrix is ​​constructed according to the connection relationship between the devices in the acquisition processing unit (1) to represent the connection status between the nodes in the graph, and the device power, operating time and device type of each node are used as a feature matrix; Initialize the dimension of the convolution kernel, the convolution kernel belongs to the graph convolutional neural network itself, the convolution kernel dimension matches the dimension of the node feature matrix, and the initialization weight of the convolution kernel is randomly assigned based on the distribution of node features; The convolution kernel performs a sliding convolution operation on the graph structure data. For each node, the convolution kernel determines its neighbor nodes based on the adjacency matrix, and then multiplies its own weight with the eigenvalues ​​of the central node and its neighbor nodes and sums them up to aggregate local features, obtain the impact of the relationship between devices on the load, and dynamically adjust the calculation weight of the convolution kernel on the node according to the contribution of the node during the convolution process.

5. A building energy consumption management system for analyzing power grid load according to claim 4, characterized in that: The graph convolutional neural network in the load prediction module (21) extracts the relationship features between devices for load prediction during the convolution operation process, which is specifically as follows: For the first layer of the graph convolutional neural network, the convolution kernel function is determined according to the number of nodes, connection density and initial features of the nodes in the initial graph structure. If the number of nodes is N, the connection density is D, the feature dimension is F, the convolution kernel size is set to k1×k1, the step size is set to s1, the padding method is p1 and the number of convolution kernels is set to c1; For each node, the convolution kernel determines its neighboring nodes based on the adjacency matrix, multiplies the convolution kernel weight with the eigenvalues ​​of the central node and its neighboring nodes, and sums them up, completing local feature aggregation and extracting the primary features of the relationship between the node and its neighboring nodes. After the first layer of convolution, the nodes are clustered according to the extracted features. In the second layer, if the output feature dimension of the first layer is F1, the convolution kernel size is k2×k2, (k2>k1), the step size is s2, and the padding method is p2, the convolution operation is performed based on the adjusted parameters, the local relationship extracted by the first layer is integrated, the primary node features are updated, and the intermediate node features are obtained. At the same time, the skip connection is introduced in the second layer of convolution to directly transfer the feature information of the first half of the first layer to the second layer; For the third layer, if the output feature dimension of the second layer is F2, the convolution kernel size is set to k3×k3, (k3>k2), the step size is set to s3, the padding method is set to p3, and the convolution kernel weight is updated according to the output feature of the second layer using the adaptive learning rate adjustment algorithm. At the same time, the number of convolution kernels is increased to c3; The adjusted parameters are used for convolution to obtain the load relationship characteristics between the devices in the graph, and load prediction is performed based on the load relationship characteristics between the devices in the graph.

6. A building energy consumption management system for analyzing power grid load according to claim 1, characterized in that: The operation of the attention mechanism in the analysis and tracing module (22) when weighting the device operation data is performed is as follows: Perform linear transformation on the equipment operation data, set the equipment operation data matrix X, and multiply the equipment operation data matrix by the preset weight matrix W q , W k and W v , generate query vector, key vector and value vector respectively; The similarity between the query vector and the key vector is calculated to form a similarity matrix, and the similarity matrix is ​​normalized to obtain weights. The weights are used to perform weighted summation on the value vectors, and the weighted summation is used to assign weights to the device operation data.

7. A building energy consumption management system for analyzing power grid load according to claim 1, characterized in that: The operation of the hidden Markov model in the decomposition and tracing module (22) when modeling the composite mode under different states of the device is as follows: Extracting the power, current, voltage and operating frequency of the device from the acquisition unit (1), and extracting the mean value and the statistical characteristics of the correlation between the parameters; Obtain the equipment's past operating status records and load data from the historical database, compare and integrate the current data features with them, and identify the operating mode and status change trend; Based on the above analysis, the equipment operation conditions with the same characteristics and historical performance are classified into categories, and each category is an equipment state, including normal operation, different load levels, and fault operation states; For each equipment state, the occurrence frequency of different load value ranges in this state is counted, and a normal distribution is constructed to describe the load pattern. The parameters of the probability distribution function are determined by maximum likelihood estimation. When new equipment operation data is transmitted, the Bayesian theorem method is used to calculate the probability of occurrence in each state by comparing the new data with the characteristics of each state and the probability distribution of the load pattern, so as to determine the current state of the equipment, perform load decomposition and trace the source, and introduce prior knowledge to correct the probability calculation.

8. A building energy consumption management system for analyzing power grid load according to claim 7, characterized in that: During the operation of the hidden Markov model in the decomposition and tracing module (22), the probability distribution of the load mode in each state is updated in real time according to the newly acquired equipment operation data, as follows: When new equipment operation data is obtained, for each determined equipment state in the model, the load value under that state is re-counted, and the load pattern probability distribution under each state is updated in real time according to the value; The device was in state S at the previous moment. a , in state S at the current moment b The probability is the state transition probability. According to Bayes' theorem, after new data is acquired, these probability values ​​are recalculated, and the state transition probability is adjusted through the calculation.

9. A method for implementing a building energy management system for analyzing power grid load according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1. Extract the power, current and voltage of the electrical equipment in the building and the temperature and humidity data of the building environment, align the equipment operation data and meteorological data in time series, and construct the connection relationship data between devices into graph structure data; S2. First, input the preprocessed time series data into the long short-term memory network step by time, and capture the time dependency through the coordinated operation of the forget gate, input gate and output gate. Then, input the data processed by the long short-term memory network and the graph structure data into the graph convolutional neural network. The graph convolutional neural network constructs the adjacency matrix and feature matrix according to the device connection relationship. After adjusting the multi-layer convolution kernel parameters, the local features of the nodes are gradually extracted to the load relationship features between devices to perform load forecasting. S3. Use the attention mechanism to linearly transform the device operation data matrix and generate a vector with the preset weight matrix, calculate the similarity, normalize the weight to obtain the weight to assign the data weight, and combine the hidden Markov model to extract the device power characteristics and integrate historical data to determine the state and load mode probability distribution. When new data is passed in, use the Bayesian theorem to calculate the probability to determine the state to achieve decomposition and traceability. During operation, the load mode probability distribution and state transition probability of the state are also updated in real time; S4. Based on the predicted grid load, the time period is determined by comparing it with the peak and valley thresholds set based on the historical energy consumption data of the building and the power supply capacity of the grid. During the peak and valley periods, instructions for adjusting operating parameters are issued to the adjustable equipment. At the same time, based on the load decomposition and tracing results, the probability distribution of the equipment energy consumption pattern and the normal load pattern are compared, and fault warnings are issued for abnormal energy consumption equipment.

Images