A commercial building short-term load prediction method based on personnel occupancy information fusion

Abstract

This invention relates to a short-term load forecasting method for commercial buildings based on occupancy information fusion. The method is characterized by the following steps: First, the CEEMDAN algorithm is used to forward decompose the original load data into multiple IMFs (Integrated Motion Factors). After removing high-frequency noise components, the remaining components are reconstructed and fused to obtain clean load data. Then, a DWedRVFL neural network model is constructed. This model uses the latest prediction accuracy to capture the dynamic changes in the load sequence and constructs a diversity index to utilize the multi-scale features of different output layers. Furthermore, a ranking strategy is constructed to integrate the contributions of the latest accuracy and diversity, while avoiding the impact of outlier predictions on the combined forecast. Finally, historical load, weather conditions, and occupancy data are combined into a connection matrix and fed into the DWedRVFL neural network model for learning, thereby improving the accuracy and stability of building load forecasting. This method is beneficial for improving the accuracy of short-term load forecasting for commercial buildings.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and specifically to a short-term load forecasting method for commercial buildings based on the fusion of occupancy information. Background Technology

[0002] Commercial buildings account for a large proportion of global energy consumption and carbon dioxide emissions, and this proportion is continuously increasing. Commercial buildings equipped with Building Energy Management Systems (BEMS) are well-suited for optimizing and reducing building energy consumption. BEMS can utilize existing energy resources in a more intelligent and refined manner, improving building energy efficiency and implementing the "carbon peaking" and "carbon neutrality" strategies. Accurate short-term load forecasting for commercial buildings is the foundation for developing BEMS energy dispatching, energy-saving control, and demand response strategies.

[0003] Commercial building loads are not only closely related to seasonality, weather, and climate, but are also significantly influenced by the random occupancy behavior of people within the building. The presence of people causes temperature variations in the indoor environment, thus affecting the power consumption of air conditioning systems; occupancy of energy-consuming equipment such as air conditioners, lighting fixtures, and computers causes fluctuations in building load. Therefore, short-term load forecasting modeling for commercial buildings must consider occupancy information to achieve higher accuracy in forecasting performance.

[0004] Chinese patent application number 202410429621.5 discloses a building load forecasting method and related system based on transfer learning. This method first acquires source and target domain data through a data acquisition module, and then fills in missing data in the target domain data based on the source domain data. A feature extractor is used to extract and reconstruct the temporal features in the load data. A regression predictor is used to predict the load data in the source domain. The trained model is then transferred and retrained and evaluated on the target domain test data. However, this method does not consider the impact of occupancy behavior on building load.

[0005] Chinese patent application number 201910580594.0 discloses a renewable energy system based on indoor occupant behavior and its load forecasting method. This method obtains indoor occupant information from a human behavior module, then uses this occupant information, meteorological data from the meteorological bureau, and historical load data as input parameters for a load forecasting module. Next, it establishes a building load forecasting model based on a BP neural network to predict the building's load at a future time. However, the occupant occupancy data in this method is generated based on a simulation model and cannot effectively detect occupancy information in real-world scenarios, thus affecting the accuracy of load forecasting. Summary of the Invention

[0006] The purpose of the present invention is to provide a short-term load forecasting method for commercial buildings based on the fusion of occupancy information, which is beneficial to improving the accuracy of short-term load forecasting for commercial buildings.

[0007] To achieve the above purpose, the technical solution adopted by the present invention is as follows: First, a wireless sensing solution is constructed to accurately, low-costly, and non-invasively obtain occupancy information using the existing wireless devices in the BEMS of commercial buildings, and to reflect the relationship between personnel occupancy and building energy consumption; Second, to effectively integrate the personnel occupancy information of multiple rooms in the building, an improved stacked sparse autoencoder EKSSAE is constructed; expert knowledge is embedded in the unsupervised training process of EKSSAE to guide the model to prioritize important information sources to improve the quality of the fused data; Finally, the building-level occupancy information, temperature, and calendar indication are combined with historical load information to form a connection matrix and fed into a random vector functional link neural network RVFL for learning. The trained RVFL network can accurately predict the building's electricity load in real time.

[0008] Furthermore, for intelligent commercial buildings equipped with BEMS, a LoRa sensing device is installed in each room of the building to transmit the energy consumption and status data of energy-consuming devices; when someone is present in the office, the LoRa channel state changes; the occupancy information is reflected by the received signal strength indicator RSSI generated by the data transmission of the LoRa device; the RSSI signals of all rooms in the collected commercial building are expressed as M = [m1, m2, …, m n T , where n is the number of signal data points.

[0009] Furthermore, in order to keep the sampling frequency of the LoRa signal consistent with the load data and at the same time extract the important occupancy features in the signal, a time aggregation-based data alignment method is used to process the LoRa signal; let the load data sampling interval be H, and H < N, then the following formula is used to perform time aggregation processing on h LoRa signal data within the H interval:

[0010]

[0011] Aggregate the signals within all H intervals according to the above formula, and then aggregate the signals of each room in this way to obtain the final aggregated signal S contains the personnel occupancy information in the room, N represents the number of aggregated LoRa signal data, and m is the number of rooms.

[0012] ​Furthermore, wireless sensing devices are deployed in different rooms of the smart building; it is necessary to fuse the occupancy data S from different rooms into comprehensive, high-quality building-level occupancy information to improve the accuracy of building load prediction; considering the differences in the importance of information sources in different rooms, an improved stacked sparse autoencoder EKSSAE based on expert knowledge is constructed to perform the occupancy information fusion process; firstly, the importance weights of the information sources are designed according to expert knowledge:

[0013]

[0014] In the formula, n c and n p These represent the room's monthly electricity consumption and daily number of office workers, respectively; c is the Spearman correlation coefficient between the LoRa signal and the building's load; n c and n p The values ​​are normalized to ensure dimensional consistency. The normalization formula is:

[0015]

[0016] Where, n c,min and n c,max Representing n respectively c The minimum and maximum values ​​in n c,nor The data is after normalization; This represents an importance weight vector based on expert knowledge; the larger the weight value, the higher the importance.

[0017] Furthermore, aggregation signal First, the input is fed into the activation layer of EKSSAE for calculation, and the activation output of the j-th neuron in the activation layer is calculated. Represented as:

[0018] φ=f(W c x+b c ),

[0019] Among them, W c and b c Let represent the weights and biases of the activation layer, respectively; the penalty term constructed in the activation layer based on the KL divergence coefficient and the expert knowledge matrix ψ is:

[0020]

[0021] Where μ represents the penalty factor, and ψ j Represents the average activation value and target activation value of the j-th neuron in the activation layer; importance penalty term J isThe actual activation probability of neurons in the expected activation layer matches the expected probability, thus EKSSAE can guide the network to assign higher target activation values ​​to important features; the activation layer does not use a fully connected structure; J is Calculations are performed only in the first AE unit.

[0022] Furthermore, the output features of the activation layer Continue learning in subsequent layers; the encoding and decoding process of the first layer in EKSSAE is represented as follows:

[0023] h (1) =f(W e φ+b e )

[0024]

[0025] In the formula, h (1) and Let represent the hidden layer features and reconstructed output of the first layer of EKSSAE, respectively; the sparse penalty terms for the subsequent k stacked autoencoder units are shown below:

[0026]

[0027]

[0028] In the formula, η is the sparsity penalty factor, which controls the degree of penalty for sparsity constraints; ρ (k) It is the target activation probability value of the k-th layer; KL divergence can measure The difference between them, when When the KL divergence value is at its minimum.

[0029] Furthermore, since the impact of occupancy information reflected by LoRa RSSI in different rooms on building load varies, importance weights are embedded in the MSE loss function:

[0030]

[0031] Among them, J im The improved MSE loss function, based on importance, amplifies the contribution of important features to the overall loss, enabling the network to focus more on this crucial information. Furthermore, due to the addition of an extra activation layer, the regularization constraint expression is written as:

[0032]

[0033] Among them, J ir It is an improved regularization penalty term, W. c This represents the connection weights between the output layer and the activation layer. Let represent the encoder weights and decoder weights of the k-th AE unit, respectively; therefore, the training objective function of EKSSAE can be written as:

[0034] ISSAE loss =J im +J s +J is +J ir

[0035] SSAE is trained layer by layer in an unsupervised manner, with the training objective being to minimize ISSAE. loss To find the optimal set of parameters; ultimately, EKSSAE retains only the encoder part, and the hidden layer feature output generated by the last autoencoder unit is used as the fusion signal, which is represented as:

[0036] Furthermore, construct datasets for training and testing; given an input, X L Represents the historical load sequence, the remaining inputs X R X W and X C These represent historical LoRa signal sequences, weather forecast variables, and calendar variables, respectively; the goal is to establish the following mapping relationship:

[0037]

[0038] In the formula, Let 'e' be the predicted load value for the next time step, and 'e' be the error; the rolling window 'w' determines the amount of historical data used; where the historical load sequence and the personnel occupancy data sequence are respectively... And X W ,X C It is a scalar; [X] L ,X R ,X W ,X C The connection matrix X is formed as the input to the prediction model, and the output is the predicted load value for the next time step.

[0039] Furthermore, the input data is represented as Where N and m are the number of input data samples and the number of features; P is defined as the number of neurons in the enhancement layer; the features of the enhancement layer are represented as... It is calculated by the following formula:

[0040] H = g(XB1)

[0041] Where g(·) represents a nonlinear activation function, such as sigmoid, tanh, or ReLU, etc. The connection weights are randomly generated for the enhancement layer; the enhancement features and input data are combined into a sequence matrix [H,X] and fed into the linear output layer; then the output of RVFL is... for:

[0042]

[0043] Where β represents the weights connecting to the output layer; since the training process of RVFL only involves the calculation of the output weights β, the loss function of RVFL is expressed as:

[0044] Loss RVFL =||[[H,X]β-|Y|| 2 +λ||β|| 2

[0045] Where λ is the regularization parameter; generally, β is calculated analytically using the Moore-Penrose pseudo-inverse method or ridge regression method, expressed as:

[0046]

[0047] in, It is a combination matrix, where I represents the identity matrix; the training batch size, learning rate, activation function, and number of hidden layer neurons of the RVFL network are set and trained, with the training objective being the loss. RVFL Minimize; at the end of training, the RVFL network outputs the optimal output layer weight parameter β; subsequently, the test set data is fed into the RVFL for prediction.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1) Random behavior of people in buildings is a major factor affecting the accuracy of building load forecasting; therefore, it is necessary to incorporate occupancy information into the building load forecasting model. However, existing technologies have not addressed the issues of low accuracy, high cost, and infringement of user privacy in acquiring occupancy information. To address these issues, this invention proposes a wireless sensing-based building occupancy information collection system that utilizes pre-existing LoRa devices in the BEMS system to sense occupancy. This solution can acquire occupancy data in real time, accurately, and at low cost without infringing on occupancy privacy.

[0050] 2) Occupancy information sources for each room in a smart building can only reflect the occupancy status of a specific area. Existing commercial building prediction methods do not consider the fusion of occupancy information from different areas. To obtain high-quality building-level occupancy information, this invention designs an occupancy information fusion framework based on an improved stacked sparse autoencoder (EKSSAE). EKSSAE introduces domain expert knowledge to improve the network's focus on important information, thereby obtaining high-quality building occupancy information.

[0051] 3) An input connection matrix is ​​constructed based on occupancy information, weather, calendar, and historical load data, and then fed into RVFL for training. Unlike traditional feedforward neural networks, RVFL has a direct link from the input layer to the output layer and uses randomization techniques to generate initial network parameters. These characteristics ensure the learning efficiency and generalization ability of RVFL, thus enabling high-precision short-term building load forecasting.

[0052] 4) Compared with the patent "A building load forecasting method and related system based on transfer learning", this invention considers the impact of occupancy on load changes in the load forecasting modeling stage, thus effectively explaining the random fluctuations in building load; compared with the patent "A renewable energy system based on indoor human behavior and its load forecasting method", the wireless sensing scheme developed in this invention can obtain occupancy status in real time and accurately, and obtain high-quality building-level occupancy information through the designed data fusion framework. The novel input system based on occupancy information significantly improves the accuracy of short-term building load forecasting. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the method implementation of an embodiment of the present invention;

[0054] Figure 2 This is an architecture diagram of the RVFL prediction network in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of a building occupancy information collection system based on wireless sensing in an embodiment of the present invention;

[0056] Figure 4 This is an example LoRa RSSI data image collected from room B in this embodiment of the invention;

[0057] Figure 5 This is an aggregated signal diagram of 26 rooms in an embodiment of the present invention;

[0058] Figure 6 This is a fused signal diagram in an embodiment of the present invention;

[0059] Figure 7 This is a prediction result diagram of an embodiment of the present invention. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0062] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0063] Random occupancy behavior affects the accuracy of short-term load forecasting for commercial buildings. To address this issue, this embodiment proposes a short-term load forecasting method for commercial buildings based on occupancy information fusion, the implementation process of which is as follows: Figure 1 As shown, firstly, a wireless sensing scheme was developed to accurately, cost-effectively, and non-intrusively acquire occupancy information using existing wireless devices within the Building Energy Management System (BEMS) of commercial buildings, reflecting the relationship between occupancy and building energy consumption. Secondly, to effectively integrate occupancy information from multiple rooms within the building, an improved stacked sparse auto-encoder based on expert knowledge (EKSSAE) was constructed. Expert knowledge was embedded into the unsupervised training process of EKSSAE, guiding the model to prioritize important information sources to improve the quality of the fused data. Finally, a connection matrix composed of building-level occupancy information, temperature, calendar indicators, and historical load information was fed into a random vector functional link (RVFL) neural network for learning. The trained RVFL network can accurately predict the building's power load in real time. Figure 2 This is the architecture diagram of the RVFL network in this embodiment.

[0064] This method applies to smart commercial buildings with two or more floors equipped with BEMS (Building Automation System). The primary function of commercial buildings is to provide office space for business professionals. Each room in a smart commercial building is equipped with a LoRa (Long Range) sensor, which is part of BEMS and used to transmit energy consumption and status data of the room's energy-consuming devices. This method is not applicable to commercial buildings that have not undergone intelligent renovations or are not equipped with BEMS.

[0065] The short-term load forecasting method for commercial buildings based on the fusion of personnel occupancy information is specifically implemented according to the following steps:

[0066] Step 1: The object of this invention's research is an intelligent commercial building equipped with a BEMS. A LoRa (Long Range) sensing device is installed in each room of the building to transmit the energy consumption and status data of energy-consuming devices. When someone is occupying an office, the LoRa channel state will change. The occupancy information is reflected by the received signal strength indicator RSSI (received signal strength indicator) generated by the LoRa device data transmission, and this data can be conveniently obtained through the database. The RSSI signals of all rooms in the collected commercial building are represented as M = [m1, m2, …, m n T , where n is the number of signal data points.

[0067] Step 2: In order to keep the sampling frequency of the LoRa signal consistent with the load data and extract the important occupancy features in the signal, a data alignment method based on time aggregation is used to process the LoRa signal. Let the sampling interval of the load data be H, and H < N, then the time aggregation processing of h LoRa signal data within the H interval is carried out according to the following formula:

[0068]

[0069] Aggregate the signals within all H intervals according to the above formula, and then aggregate the signals of each room in this way to obtain the final aggregated signal S contains the personnel occupancy information in the room, N represents the number of aggregated LoRa signal data, and m is the number of rooms.

[0070] Step 3: Wireless sensing devices are deployed in different rooms of the intelligent building. It is necessary to fuse the occupancy data S of different rooms into comprehensive and high-quality building-level occupancy information to improve the accuracy of building load forecasting. Considering the differences in the importance of different room information sources, an improved stacked sparse autoencoder EKSSAE based on expert knowledge is constructed to perform the fusion process of occupancy information. First, design the importance weight of the information source according to expert knowledge:

[0071]

[0072] In the formula, n c and n p respectively represent the monthly electricity consumption and the daily office population of this room; c is the Spearman correlation coefficient between the LoRa signal and the building load. n c and n​p The values ​​are normalized to ensure dimensional consistency. The normalization formula is:

[0073]

[0074] Where, n c,min and n c,max Representing n respectively c The minimum and maximum values ​​in n c,nor This is the normalized data. This represents an importance weight vector based on expert knowledge; the larger the weight value, the higher the importance.

[0075] Step 4: Aggregate Signals First, the input is fed into the activation layer of EKSSAE for calculation, and the activation output of the j-th neuron in the activation layer is calculated. It can be represented as:

[0076] φ=f(W c x+b c ),

[0077] Among them, W c and b c Let represent the weights and biases of the activation layer, respectively. A penalty term based on the KL divergence coefficient and the expert knowledge matrix ψ is constructed in the activation layer as follows:

[0078]

[0079] Where μ represents the penalty factor, and ψ j This represents the average activation value and the target activation value of the j-th neuron in the activation layer. Importance penalty term J. is The expected activation probability of neurons in the activation layer matches the expected probability, thus EKSSAE guides the network to assign higher target activation values ​​to important features. Note that the activation layer does not use a fully connected structure because the input data loses its original feature information after computation through a fully connected layer, and the importance information no longer corresponds to the original input. Therefore, J is Calculations are performed only in the first AE unit.

[0080] Step 5: Output features of the activation layer Continue learning in subsequent layers. Therefore, the encoding and decoding process of the first layer in EKSSAE can be represented as:

[0081] h (1) =f(W e φ+b e )

[0082]

[0083] In the formula, h (1) and Let represent the hidden layer features and reconstructed output of the first layer of EKSSAE, respectively. The sparse penalty terms for the subsequent k stacked autoencoder units are shown below:

[0084]

[0085] In the formula, η is the sparsity penalty factor, which controls the degree of penalty for sparsity constraints; ρ (k) This is the target activation probability value of the k-th layer. KL divergence can be measured... The difference between them, when When the KL divergence value is at its minimum.

[0086] Step 6: Since the impact of occupancy information reflected by LoRa RSSI in different rooms on building load varies, it is necessary to embed importance weights into the MSE loss function:

[0087]

[0088] Among them, J im The improved MSE loss function, based on importance, amplifies the contribution of important features to the overall loss, enabling the network to focus more on this crucial information. Furthermore, due to the addition of an extra activation layer, the regularization constraint expression can be written as:

[0089]

[0090] Among them, J ir It is an improved regularization penalty term, W. c This represents the connection weights between the output layer and the activation layer. and Let represent the encoder weights and decoder weights of the k-th AE unit, respectively. Therefore, the training objective function of EKSSAE can be written as:

[0091] ISSAE loss =J im +J s +J is +J ir

[0092] SSAE is trained layer by layer in an unsupervised manner, with the training objective being to minimize ISSAE. loss To find the optimal set of parameters. Ultimately, EKSSAE retains only the encoder part, and the hidden layer feature output generated by the last autoencoder unit is used as the fusion signal, which can be expressed as:

[0093] Step 7: Construct the dataset for training and testing. Given an input, X L Represents the historical load sequence, the remaining inputs X R X W and X C These represent historical LoRa signal sequences, weather forecast variables, and calendar variables, respectively. The goal is to establish the following mapping relationship:

[0094]

[0095] In the formula, Let be the predicted load value for the next time step, and 'e' be the error. The rolling window 'w' determines the amount of historical data used. The historical load series and the personnel occupancy data series are respectively... And X W ,X C It is a scalar. [X] L ,X R ,X W ,X C The connection matrix X is formed as the input to the prediction model, and the output is the predicted load value for the next time step.

[0096] Step 8: Input data is represented as Where N and m are the number of input data samples and the number of features, respectively. P is defined as the number of neurons in the enhancement layer. The features of the enhancement layer are represented as follows: It can be calculated by the following formula:

[0097] H = g(XB1)

[0098] Where g(·) represents a nonlinear activation function, such as sigmoid, tanh, or ReLU, etc. The connection weights are randomly generated for the enhancement layer. The enhancement features and input data are combined into a sequence matrix [H,X] and fed into the linear output layer. The output of the RVFL is then... for:

[0099]

[0100] Where β represents the weights connecting to the output layer. Since the training process of RVFL only involves calculating the output weights β, the loss function of RVFL can be expressed as:

[0101] Loss RVFL =||[[H,X]β-|Y|| 2 +λ||β|| 2

[0102] Where λ is the regularization parameter. Generally, β can be calculated analytically using the Moore-Penrose pseudoinverse or ridge regression method, expressed as:

[0103]

[0104] in, It is a combination matrix, where I represents the identity matrix. The training batch size, learning rate, activation function, and number of hidden layer neurons of the RVFL network are set and trained, with the training objective being the loss... RVFL Minimize. At the end of training, the RVFL network outputs the optimal output layer weight parameters β. Subsequently, test set data is fed into the RVFL for prediction.

[0105] The specific implementation of this invention will be described in detail below with reference to data examples. The specific steps are as follows:

[0106] Step 1: The data studied in this paper comes from a commercial building with three floors and 26 rooms in Fuzhou Software Park. The designed building occupancy information collection system based on wireless sensing is as follows: Figure 3 As shown. In a smart commercial building equipped with BEMS, each room is equipped with a LoRa module for transmitting energy consumption and status data of energy-consuming devices. This data is transmitted from the transmitter to the receiver, and the corresponding LoRa signal is obtained based on the device identifier of each room. When someone is present in the office, the LoRa RSSI channel status changes, reflecting occupancy information. LoRa RSSI data can be easily accessed through a database. The designed occupancy information collection system was used to collect LoRa RSSI data from all 26 rooms of this commercial building during weekdays in July 2023. Figure 4 Example LoRa RSSI data collected for office B.

[0107] Step 2: The data size of the raw LoRa RSSI signal data matrix M collected from all 26 rooms is 26 × 7770 = 202020. To maintain consistency between the LoRa signal and load data sampling frequencies and to extract important occupancy features from the signals, a time-aggregation-based data alignment method is used to process the LoRa signals. Assuming the load data sampling interval is H = 30 minutes, time aggregation is performed on h LoRa signal data within interval H. The value of h is related to the number of LoRa signals within H. The number of signal data within the first H interval is equal to 6. Therefore, the aggregated signal within the first H interval is:

[0108]

[0109] The aggregated signal within each H is calculated using the above formula; then the signals from each room are aggregated in this manner, ultimately yielding the aggregated signal sequence. Where N = 864, m = 26; then the aggregated signal S of all 26 rooms is as follows: Figure 5 As shown.

[0110] Step 3: Wireless sensing devices are deployed in different rooms of the smart commercial building. It is necessary to fuse the occupancy data from different rooms into comprehensive, high-quality building-level occupancy information to improve the accuracy of building load prediction. Considering the differences in the importance of information sources in different rooms, an improved stacked sparse autoencoder (EKSSAE) is designed to perform the occupancy information fusion process. Taking office B as an example, the importance weights of the information sources are designed based on expert knowledge:

[0111]

[0112] In the formula, n c =4.22 and n p = 15 represents the monthly electricity consumption and average daily number of office workers in room B, respectively; c is the Spearman correlation coefficient between the LoRa signal in room A and the building load, which is calculated to be 0.452. c and n p The values ​​are normalized to ensure dimensional consistency. The importance weights based on expert knowledge for all 26 rooms are calculated using the formula above: ψ = [9.25e-03, 5.95e-01, 4.98e-01, 2.06e-02, 1.20e-02, 2.95e-01, 7.50e-05, 1.54e-02, 2.80e-04, 8.75e-03, 8.87742386e-05, 1.04e-04, 2. [29e-04,7.95e-02,1.11e-04,2.22e-02,4.96e-05,7.55e-05,1.61e-02,2.26e-04,2.25e-04,2.89e-04,9.99e-05,9.41e-05,1.99e-04,3.484e-04], where the importance value of room B is 5.95e-01.

[0113] Step 4: Set the learning rate and batch size of the EKSSAE network to 0.001 and 32, respectively, and set the number of training iterations to 500. The specific training parameters are shown in Table 1.

[0114] Table 1

[0115]

[0116] First, the aggregated wireless signal sequence data from the 26 rooms were... The input is fed into the activation layer of EKSSAE for calculation, and then the activation output is activated. It can be represented as:

[0117] φ=f(W c x+b c ),

[0118] Among them, W c and b c Let represent the weights and biases of the activation layer, respectively. The penalty term based on the KL divergence coefficient is designed in the activation layer as follows:

[0119]

[0120] Here, μ = 1 represents the penalty factor. Taking the second neuron in the activation layer as an example, then... and ψ j =0.595 represents the average activation value and target activation value of the second neuron in the activation layer.

[0121] Step 4: Features output from the activation layer Continue learning in subsequent layers. Therefore, the encoding and decoding process of the first AE unit in EKSSAE can be represented as:

[0122] h (1) =f(W e φ+b e ),

[0123]

[0124] In the formula, h (1) and These represent the hidden layer features and reconstructed output of the first layer of EKSSAE, respectively. Taking k=3 as an example, the sparse penalty term of its autoencoder unit is as follows:

[0125]

[0126] In the formula, η = 1 is the sparsity penalty factor, which controls the degree of penalty for sparsity constraints; ρ (k) The target activation probability value for layer 3 is set to 0.05.

[0127] Step 4: Since the impact of occupancy information reflected by LoRa RSSI in different rooms on building load varies, it is necessary to embed importance weights into the MSE loss function:

[0128]

[0129] Among them, J imThe improved MSE loss function, based on importance, amplifies the contribution of important features to the overall loss, enabling the network to focus more on this crucial information. Furthermore, due to the addition of an extra activation layer, the regularization constraint expression can be written as:

[0130]

[0131] Among them, J ir It is an improved regularization penalty term, W. c This represents the connection weights between the output layer and the activation layer. Let represent the encoder weights and decoder weights of the k-th AE unit, respectively, and γ = 0.001 be the regularization coefficient. Therefore, the training objective function of EKSSAE can be written as:

[0132] ISSAE loss =J im +J s +J is +J ir .

[0133] SSAE is trained layer by layer in an unsupervised manner, with the training objective being to minimize ISSAE. loss To find the optimal set of parameters. Ultimately, EKSSAE retains only the encoder part, and the hidden layer feature output generated by the last autoencoder unit is used as the fusion signal, which can be represented as: Its like Figure 6 As shown.

[0134] Step 5: Construct the input dataset. L Represents the historical load sequence, the remaining inputs X R X W and X C These represent historical LoRa signal sequences, weather forecasts (temperature), and calendar variables (hours), respectively. The goal is to establish the following mapping relationship:

[0135]

[0136] In the formula, Let be the predicted value for the next moment, and 'e' be the error. The rolling window w = 48. The historical load sequence and personnel occupancy data are respectively... And X W ,X C It is a scalar. [X] L ,X R ,X W ,X C Forming a connection matrix As input to the prediction model, the number of samples in the input matrix is ​​N = 864, and the number of features is m = 48 + 48 + 2 = 98; that is, the dimension of each input vector in the dataset is 98, and the output is the load value at the next time step.

[0137] Step 6: Set the training batch size, learning rate, and activation function parameters for the RVFL network to 32, 0.01, and sigmoid, respectively. Set the training stopping condition to 500 generations. Divide the training, validation, and test sets into a 7:1:2 ratio, resulting in a training data size of 605 × 4 = 2420. Set P = 160 to the number of neurons in the enhancement layer. The RVFL model parameters are shown in Table 2. The features of the enhancement layer are represented as follows: It can be calculated by the following formula:

[0138] H = g(XB1),

[0139] Where g(·) represents the sigmoid activation function, The connection weights are randomly generated for the enhancement layer.

[0140] Table 2

[0141]

[0142] The enhanced features and input data are combined into a sequence matrix and fed into the linear output layer. The output of the RVFL is then... for:

[0143]

[0144] Where β represents the weights connecting to the output layer. Since the training process of RVFL only involves calculating the output weights β, the loss function of RVFL can be expressed as:

[0145] Loss RVFL =||[[H,X]β-|Y|| 2 +λ||β|| 2 ,

[0146] Where λ = 0.1 is the regularization parameter. Generally, β can be calculated analytically using the Moore-Penrose pseudoinverse or ridge regression method, expressed as:

[0147]

[0148] in, It is a combined matrix, where I represents a single matrix, and the training objective is Loss. RVFL Minimize. At the end of training, the RVFL network outputs the optimal output layer weights β. Subsequently, test set data is fed into the RVFL for prediction, and the prediction results are as follows... Figure 7 As shown.

[0149] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0150] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0151] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0152] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0153] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for short-term load forecasting of commercial buildings based on occupancy information fusion, characterized in that, First, a wireless sensing solution is constructed to accurately, cost-effectively, and non-invasively obtain occupancy information using existing wireless devices within the BEMS of commercial buildings, and to reflect the relationship between personnel occupancy and building energy consumption. Second, in order to effectively integrate the personnel occupancy information of multiple rooms within the building, an improved stacked sparse autoencoder EKSSAE is constructed. Expert knowledge is embedded in the unsupervised training process of EKSSAE to guide the model to prioritize important information sources to improve the quality of the fused data. Finally, the building-level occupancy information, temperature, and calendar indicating compliance with historical load information are combined to form a connection matrix and fed into the random vector functional link neural network RVFL for learning. The trained RVFL network can accurately predict the building's electrical load in real time. Wireless sensing devices are deployed in different rooms of the intelligent building. It is necessary to fuse the occupancy data S from different rooms into comprehensive and high-quality building-level occupancy information to improve the accuracy of building load prediction. Considering the differences in the importance of information sources in different rooms, an improved stacked sparse autoencoder EKSSAE based on expert knowledge is constructed to perform the fusion process of occupancy information. First, the importance weights of information sources are designed according to expert knowledge: In the formula, and These represent the monthly electricity consumption and daily number of office workers in the room, respectively; c is the Spearman correlation coefficient between the LoRa signal and the building load. and The values ​​are normalized to ensure dimensional consistency. The normalization formula is: in, and They represent The minimum and maximum values ​​in The data is after normalization; This represents an importance weight vector based on expert knowledge; a larger weight value indicates higher importance. Aggregate signal S First, the input is fed into the activation layer of EKSSAE for calculation, and the activation output of the j-th neuron in the activation layer is calculated. Represented as: in, and These represent the weights and biases of the activation layer, respectively; a matrix based on KL divergence coefficients and expert knowledge is constructed in the activation layer. The penalties are as follows: Where μ represents the penalty factor, and Represents the average activation value and target activation value of the j-th neuron in the activation layer; importance penalty term. The actual activation probability of neurons in the expected activation layer is consistent with the expected probability, thus EKSSAE can guide the network to assign higher target activation values ​​to important features; the activation layer does not use a fully connected structure; Calculated only in the first AE unit; The output features φ of the activation layer continue to be learned in subsequent layers. The encoding and decoding process of the first layer in EKSSAE is expressed as: In the formula, and Let represent the hidden layer features and reconstructed output of the first layer of EKSSAE, respectively; then the sparse penalty terms for the k stacked autoencoder units are as follows: In the formula, η is the sparsity penalty factor, which controls the degree of penalty for sparsity constraints; ρ (k) It is the target activation probability value of the k-th layer; KL divergence can measure and The difference between them, when = When the KL divergence value is at its minimum. 2.The commercial building short-term load forecasting method based on personnel occupancy information fusion according to claim 1, wherein, For smart commercial buildings equipped with BEMS, each room in the building is equipped with a LoRa sensor to transmit energy consumption and status data of power-consuming devices. When an office is occupied, the LoRa channel status changes. Occupancy information is reflected by the Received Signal Strength Index (RSSI) generated by the LoRa device's data transmission. The collected RSSI signals from all rooms in the commercial building are represented as follows: , where n is the number of signal data points.

3. The method for short-term load forecasting of commercial buildings based on occupancy information fusion according to claim 2, characterized in that, In order to keep the sampling frequency of the LoRa signal consistent with the load data and at the same time extract important occupancy features in the signal, a time aggregation-based data alignment method is used to process the LoRa signal. Let the sampling interval of the load data be H, and H < N. Then, the time aggregation process of h LoRa signal data within the H interval is performed according to the following formula: ×100% The signals within all H intervals are aggregated using the above formula, and then the signals from each room are aggregated in the same way to obtain the final aggregated signal S. S contains information on the occupancy of people in the room, N represents the number of aggregated LoRa signal data, and m is the number of rooms.

4. The short-term load forecasting method for commercial buildings based on personnel occupancy information fusion according to claim 1, characterized in that, Since the occupancy information reflected by the LoRa RSSI in different rooms has different impacts on the building load, importance weights are embedded in the MSE loss function: in, The improved MSE loss function, based on importance, amplifies the contribution of important features to the overall loss, enabling the network to focus more on this crucial information. Furthermore, due to the addition of an extra activation layer, the regularization constraint expression is written as: in, It is an improved regularization penalty term. This represents the connection weights between the output layer and the activation layer. and Let represent the encoder weights and decoder weights of the k-th AE unit, respectively; therefore, the training objective function of EKSSAE can be written as: SSAE is trained layer by layer in an unsupervised manner, with the training objective being to minimize To find the optimal set of parameters; ultimately, EKSSAE retains only the encoder part, and the hidden layer feature output generated by the last autoencoder unit is used as the fusion signal, which is represented as: .

5. The method for short-term load forecasting of commercial buildings based on occupancy information fusion according to claim 4, characterized in that, Construct a dataset for training and testing; given an input, Represents the historical load sequence, the rest of the input , and Let these represent the personnel occupancy data sequence, weather forecast variable, and calendar variable, respectively; the goal is to establish the following mapping relationship: In the formula, Let 'e' be the predicted load value for the next time step, and 'e' be the error; the rolling window 'w' determines the amount of historical data used; where the historical load sequence and the personnel occupancy data sequence are respectively... and ,and It is a scalar; [ Forming a connection matrix As input to the prediction model, the output is the predicted load value for the next time step.

6. The short-term load forecasting method for commercial buildings based on personnel occupancy information fusion according to claim 5, characterized in that, Input data is represented as Where N' and m' are the number of samples and features of the input data; P is defined as the number of neurons in the enhancement layer; the features of the enhancement layer are represented as... It is calculated by the following formula: Where g(⋅) represents a nonlinear activation function, The connection weights are randomly generated for the enhancement layer; the enhancement features and input data are combined into a sequence matrix. And feed it to the linear output layer; then the output of RVFL for: where β is the weight connecting the output layer. Since the training process of RVFL only involves the calculation of the output weight β, the loss function of RVFL is expressed as: where λ is the regularization parameter. The calculation of β is analytically calculated using the Moore-Penrose pseudoinverse method or the ridge regression method, and is expressed as: in, It is a combination matrix. Represent the identity matrix; set the training batch size, learning rate, activation function, and number of hidden layer neurons for the RVFL network and train it, with the training objective being... Minimize; at the end of training, the RVFL network outputs the optimal output layer weight parameter β; subsequently, the test set data is fed into the RVFL for prediction.

Images