Physical-data dual-driven subway station building load prediction method

Abstract

The present invention discloses a physical-data dual-driven subway station building load prediction method, comprising: using energy consumption simulation software to simulate and generate subway station unit area building load data under different meteorological parameters and passenger flow; decomposing the building load data into a number of intrinsic mode functions and extracting characteristic parameters; using Euclidean distance to calculate the similarity between the target subway station and multiple subway station loads; selecting a number of subway station load data with the highest similarity score according to the similarity score calculated by Euclidean distance; using the selected subway station load data to train the initial model; and predicting based on the trained prediction model. Based on the accurate prediction using the physical-data dual-driven model, the load prediction method can be transplanted to other subway stations for energy consumption analysis research, which is beneficial to the improvement of the overall energy efficiency of the subway station system.

Description

Technical Field

[0001] The present invention belongs to the field of energy consumption and energy saving of subway station buildings, and in particular to a subway station building load prediction method based on transfer learning. Background Art

[0002] Urban rail transit has become the first choice for green travel due to its safety, comfort, speed, environmental protection, large capacity, low energy consumption and low pollution. my country's subway has developed rapidly and has become one of the countries with the fastest development of urban rail transit in the world.

[0003] As of November 2024, there are 313 urban rail transit lines in operation in 54 cities across the country, with an operating mileage of 10,522.1 kilometers. Among them, there are 255 high-capacity lines such as subways, light rails, and urban rapid rails, with an operating mileage of 9,054.1 kilometers. However, with the rapid growth of the total operating mileage of rail transit, the power consumption of the system is also increasing year by year. At present, the energy consumption of the equipment system of subway stations is huge. The main reason is that the equipment selection is too large, resulting in large initial investment, and the operation stage is often operated in a "big horse pulling a small cart" mode. Even though frequency conversion technology has been widely used to save energy, the operating energy consumption of the air-conditioning system is still high due to the limitations of the system's own characteristics.

[0004] Therefore, it is necessary to start from the source of subway station equipment, improve the load calculation method, optimize equipment selection, and adopt advanced energy-saving technologies and management measures to reduce operating costs and achieve sustainable development. The energy consumption of urban rail transit systems has accounted for 3‰ of the total electricity consumption in the country, and will reach 5‰ in the future. Compared with traditional public buildings, the energy consumption level per unit area in rail transit stations is significantly higher, exceeding 200kWh / (m2•a). At present, the problem of high operating costs of rail transit is becoming increasingly prominent. High operating costs are common in rail transit projects in domestic cities. The main reason for the huge energy consumption of subway station equipment systems is that the equipment selection is much larger. Not only is the initial investment large, but the operation stage is often operated in the way of a big horse pulling a small cart. Even though frequency conversion technology has been widely used to save energy and transportation energy consumption, due to the limitations of the system's own characteristics, the operating energy consumption of the air-conditioning system is still high. Therefore, it is necessary to start from the source of subway station equipment and improve the subway station load calculation method. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a physical-data dual-driven subway station building load prediction method.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] The physical-data dual-driven subway station building load prediction method is characterized by comprising the following steps:

[0008] S1. Use energy consumption simulation software to simulate and generate unit area building load data of subway stations under different meteorological parameters and passenger flow;

[0009] S2. Decomposing the building load data generated under different meteorological parameters and passenger flow conditions by using variational mode decomposition into several intrinsic mode functions; extracting characteristic parameters from each decomposed intrinsic mode function;

[0010] S3. Based on the extracted characteristic parameters, the similarity between the target subway station and the loads of multiple subway stations is calculated using the Euclidean distance to obtain a similarity score for the load of each subway station; according to the similarity score calculated using the Euclidean distance, a similarity measurement index is determined, and several subway station load data with the highest similarity scores are selected;

[0011] S4. Pre-train the initial model using the selected subway station load data, and fine-tune the pre-trained model using a small amount of data from the target subway station to obtain a trained prediction model, so as to improve the accuracy of the prediction model in predicting the building load of the target subway station;

[0012] S5. Use the trained prediction model to perform load forecasting.

[0013] The present invention uses energy consumption simulation software such as EnergyPlus and Trnsys to create a physical model to simulate the working conditions and generate building load data under various meteorological parameters and passenger flow conditions. It combines variational mode decomposition, Euclidean distance and similarity measurement indicators to achieve high-precision prediction of subway station building loads, and can significantly improve the accuracy and robustness of predictions by utilizing the prior scientific knowledge provided by the energy consumption simulation software. A data-driven load calculation model is constructed, and based on the characteristic parameters and historical operation data of subway stations, a mapping relationship between subway station loads and key influencing factors is constructed to identify and optimize the data-driven subway station load calculation model.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] 1. Dynamic simulation using software modeling makes up for the lack of historical data and obtains more detailed dynamic air-conditioning load and energy consumption data of subway stations.

[0016] 2. Based on the characteristic analysis of subway stations and historical operation data, the mapping relationship between the total energy consumption of subway stations and influencing factors is obtained through data mining, a data-driven model is constructed, and the characteristic parameter values ​​of influencing factors are extracted to provide data support for simulation work.

[0017] 3. A load forecasting method combining LSTM long-term load forecasting and Transform short-term load forecasting is adopted to combine the advantages of long-term and short-term load forecasting to more accurately predict the load data of subway stations.

[0018] 4. The accuracy and replicability of the model are verified by using typical real site data, so that it can be transplanted to other subway stations for energy consumption analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the physical-data dual-driven subway station building load prediction method of the present invention;

[0020] Figure 2 It is a principle block diagram of the present invention using Transform to perform short-term load forecasting;

[0021] Figure 3 It is a forecasting architecture diagram of the present invention using LSTM neural network for long-term load forecasting;

[0022] Figure 4 This is the LSTM load forecast result fitting comparison chart;

[0023] Figure 5 This is the Transform load forecast result fitting comparison chart;

[0024] Figure 6 This is a comparison chart of the LSTM-Transform comprehensive load forecasting results. DETAILED DESCRIPTION

[0025] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0026] 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.

[0027] The physical-data dual-driven subway station building load prediction method provided by the present invention analyzes and predicts the energy consumption data of Suzhou subway stations based on data from previous years. The principle is to use the building load data under various meteorological parameters and passenger flow conditions generated by energy consumption simulation software such as EnergyPlus and Trnsys, combined with variational mode decomposition, Euclidean distance and similarity measurement indicators, to achieve high-precision prediction of subway station building loads, as follows:

[0028] Physical-data dual-driven subway station building load prediction method, reference Figure 1 , including the following steps:

[0029] S1. Use energy consumption simulation software such as EnergyPlus to simulate load data under different variables:

[0030] The physical model of the target subway station was created using SketchUp modeling software, and relevant parameters were entered into the energy consumption simulation software. The main parameters included: meteorological parameters (temperature, humidity, wind speed, etc.), building internal disturbance parameters (HVAC lighting equipment power), passenger flow, and equipment operation rules. After the parameters were entered, EnergyPlus and OpenStudio energy consumption simulation software were used to generate the subway station unit area building load data under different meteorological parameters and passenger flows based on historical data. The data generated by the energy consumption simulation software included but was not limited to the following variables: the number of passengers entering and leaving the subway station per unit time, building load, electric energy consumption per unit area (EEC), air conditioning energy consumption (AEC), etc., among which temperature, humidity, and wind speed were meteorological parameters, and electric energy consumption and air conditioning energy consumption were unit area building load data.

[0031] S2. Decompose the generated load data into several IMFs using VMD and propose characteristic parameters. This step includes:

[0032] S21. Variational mode decomposition (VMD) is used to decompose the building load data generated under different meteorological parameters and passenger flow conditions into several intrinsic mode functions (IMFs). The intrinsic mode function (IMF) is defined as an amplitude-frequency modulated signal. , expressed as:

[0033] in, A k ( t ) is the instantaneous amplitude of the AM / FM signal, is the instantaneous phase of the AM / FM signal, and For ease of processing, the signal is analyzed in the form of:

[0034] in, is the signal envelope, which is a non-negative function, and j is an imaginary unit.

[0035] VMD estimates the center frequency and bandwidth of each IMF by solving a constrained variational problem, which is expressed as:

[0036] in, f ( t ) is the original signal, u k ( t ) is the k IMF, ω k It is k The center frequency of the IMF is represents the convolution operation, is the unit impulse function, Indicates time t The partial derivative of .

[0037] S22. Minimize the cost function through iterative optimization to ensure that the spectrum of each modal function is concentrated near its center frequency. In the iterative process of VMD, the center frequency is iteratively updated by solving an optimization problem. ω k , can be updated by solving the following optimization problem ω k :

[0038] in, It is represented as the updated cost function after the q+1th iteration;

[0039] This optimization problem is solved numerically and its solution is expressed as:

[0040] in, u k ( ω ) yes u k ( t )’s Fourier transform;

[0041] S23. Update the modal function and center frequency until convergence.

[0042] In order to solve the above constrained variational problem, an optimization algorithm such as the alternating direction method of multipliers (ADMM) is used to alternately update the modal function u k ( t ) and center frequency ω k , at a fixed center frequency ωk In the case of , the modal function is updated by solving the following optimization problem u k ( t ):

[0043] in, is the updated mode function after q+1 iterations, α is the balance parameter of the data fidelity constraint, is the estimated value of the center frequency at the qth iteration; When the modal function uk(t) is fixed, the center frequency ωk is updated by solving the following optimization problem: .

[0044] This optimization problem can be solved by solving a ω k The minimum value of a one-variable function can be achieved by numerical methods (such as gradient descent method, Newton method, etc.), and the obtained ω k n+1 is the updated center frequency.

[0045] After each iteration, it is necessary to determine whether the algorithm has converged. The convergence condition is set based on the change in the modal function and the center frequency. If the change in the modal function and the center frequency in two consecutive iterations is less than a preset threshold (i.e., the convergence tolerance tol), the algorithm is considered to have converged and the iteration process ends. When the algorithm converges, a series of intrinsic mode functions (IMFs) and their corresponding center frequencies can be obtained. These IMFs can be used for subsequent signal analysis, processing, or reconstruction.

[0046] S24. After obtaining the IMF, extract its characteristic parameters. Common IMF characteristic parameters include:

[0047] (1) Mean: describes the average level of the signal. The calculation formula is:

[0048] in, is the nth data point in the IMF component and N is the total number of data points.

[0049] (2) Variance: describes the discreteness of the signal. Its calculation formula is:

[0050] (3) Peak value: The maximum value in the signal. Its calculation formula is:

[0051] (4) Kurtosis: describes the sharpness of the signal waveform. Its calculation formula is:

[0052] (5) Effective value: A measure of the effective energy or power of a signal. The calculation formula is:

[0053] (6) Crest factor: The ratio of the peak value to the effective value, used to describe the peak characteristics of the signal. The calculation formula is:

[0054] (7) Pulse factor: The ratio of peak value to average value, used to describe the impulse characteristics of the signal. Its calculation formula is:

[0055] (8) Waveform factor: The ratio of the effective value to the average value, used to describe the waveform characteristics of the signal. The calculation formula is:

[0056] (9) Margin factor: The ratio of peak value to root square amplitude, used to describe the dynamic range of the signal. Its calculation formula is:

[0057] S3. Use Euclidean distance to calculate the similarity between the subway station and multiple subway stations, and select several data with the highest SMI evaluation scores. This step includes:

[0058] S31. Use Euclidean distance to calculate the similarity between the IMF of the target subway station and multiple subway stations, select 3-5 characteristic parameter data (such as mean, peak factor, waveform factor) to obtain the similarity score of each subway station load; standardize the target subway station and subway station load data. X (including the characteristics of the target subway station and subway station load), the general formula for standardization is:

[0059] in, X norm is the standardized feature data, μ is the mean of the feature data, σ is the standard deviation of the feature data.

[0060] If the goal is to scale to the interval [0, 1], then the min-max normalization method can be used:

[0061] in: X min and X max are the minimum and maximum values ​​of the feature data respectively.

[0062] S32. Calculate the Euclidean distance of each characteristic data between the target subway station and the subway station load:

[0063] After the characteristic data is standardized, the target subway station (denoted as point P ) and subway station load (denoted as point Q ) between the feature data. i The calculation formula of the Euclidean distance is:

[0064] in: It is The Euclidean distance of the features, and They are points P and point Q in the The value of a feature;

[0065] S33, combining the Euclidean distances of multiple feature data to obtain an overall similarity score. After obtaining the Euclidean distances of each feature, combine the Euclidean distances to obtain a score representing the overall similarity. The overall Euclidean distance is obtained by calculating the square root of the sum of the squares of all feature Euclidean distances, and the formula is:

[0066] in: S h is the overall similarity score, z is the number of features;

[0067] S34. Conduct a comprehensive evaluation of the Similarity Measurement Index (SMI). First, take a weighted average of the Euclidean distance similarity scores of each subway station load to obtain the SMI similarity measurement index.

[0068] Let S h is the Euclidean distance similarity score between the h-th subway station load and the target subway station, and the SMI similarity measurement index is calculated:

[0069] in, m is the total number of subway station loads, It is h The weight of the subway station load is satisfied ;

[0070] If all subway station loads have equal weights (i.e. ), the formula is simplified to:

[0071] S35. Sorting the subway station loads by comparing the SMI scores of the subway station loads according to the SMI similarity measurement index. The subway station loads with higher scores have higher similarity with the target subway station. After sorting, the 100 subway station load data with the highest similarity scores are selected as the basis for subsequent processing or analysis. The specific number selected depends on the application requirements and the number of subway station loads.

[0072] S4. Use the selected data to build and train the Transformer and LSTM neural network load forecasting models:

[0073] To use the Transformer model for short-term load forecasting, you must first select a Transformer model and then choose a suitable Transformer variant (such as the original Transformer, BERT, GPT, etc.) or a custom model structure based on the specific task requirements.

[0074] Design the input layer, encoder layer, decoder layer (if necessary), and output layer of the Transformer model. In the input layer, convert the preprocessed data into a format recognizable by the model and perform position encoding to retain time series information. In the encoder layer, use components such as the multi-head attention mechanism to extract long-term dependencies in the data. In the decoder layer (if used to predict future values), use similar components to generate prediction results. In the output layer, convert the output of the decoder layer into a predicted value, and further processing (such as denormalization) may be performed to restore the original dimension.

[0075] The Transformer model is trained using the training set data. During the training process, the model parameters are updated through the back propagation algorithm to minimize the prediction error. The validation set data is used to monitor the training process and prevent overfitting.

[0076] To predict long-term load using LSTM neural network, we first forward calculate the output value of each neuron, and then reversely calculate the error term value of each neuron, including back propagation along time, that is, starting from the current time t, calculating the error term at each moment, and propagating the error term to the upper layer. Then, based on the corresponding error term, we calculate the gradient of each weight. The goal is to learn each set of parameters, including the forget gate, input gate, output gate, and the weight matrix and bias term of the unit state. The LSTM neural network training and prediction process is as follows: Figure 3 As shown, data set processing, neural network model construction, model training and prediction, and model evaluation are carried out in sequence.

[0077] LSTM uses input gate, forget gate and output gate. These gating units control the flow of information through learning, helping the LSTM network to better handle long-term dependencies.

[0078] (1) Forget Gate

[0079] The forget gate is responsible for deciding what information to discard from the cell state at the previous moment. It is based on the hidden state at the previous moment ( ) and the current input ( x t ) to make a decision.

[0080]

[0081] in: Is the output of the forget gate, indicating the degree of forgetting of each cell state value (0 means complete forgetting, 1 means complete retention) σ It is the Sigmoid function, which maps any value to the interval (0, 1); W f and b f are the weights and biases of the forget gate, which are the parameters learned during training; It means to concatenate the hidden state of the previous moment and the input of the current moment.

[0082] (2) Input gate

[0083] The input gate is responsible for deciding which new information will be stored in the cell state, which consists of two parts: one is the output of the input gate, which determines which values ​​will be updated; the other is the candidate cell state, which contains the new candidate values.

[0084] Output of the input gate:

[0085] Candidate cell states:

[0086] in: is the output of the input gate, which determines which values ​​will be updated; is the candidate cell state, which contains the new candidate value; W i , b i , W C and b C are the weights and biases of the input gate and candidate cell states.

[0087] (3) Update cell status

[0088] The cell state is updated based on the output of the forget gate and the input gate. The forget gate determines which information in the old cell state is retained, while the input gate and the candidate cell state determine which new information is added.

[0089]

[0090] in: C t is the cell state at the current moment; Indicates that the information in the old cell state is retained according to the output of the forget gate; Represents adding new information based on the output of the input gate and the candidate cell state.

[0091] (4) Output gate

[0092] The output gate is responsible for deciding which parts of the cell state will be output to the hidden state, based on the cell state and the input at the current moment.

[0093]

[0094] in: o t is the output of the output gate, which determines which parts of the cell state will be output; h t is the hidden state at the current moment and the output of the LSTM unit; W o and b o are the weight and bias of the output gate.

[0095] Model training and fine-tuning: Use the selected subway station load data to pre-train the initial model, and use a small amount of data from the target subway station to fine-tune the pre-trained model to improve the model's accuracy in predicting the building load of the target subway station; Use the selected subway station load data to pre-train the initial model, freeze some model layers, and only adjust the last layer to adapt to the target subway station data. Use the transfer learning method to train a deep neural network using subway station load data, then freeze the parameters of some of the layers, and only adjust the last few layers to adapt to the data of the target task.

[0096] Use a small amount of data from the target subway station to fine-tune the pre-trained model, update the model parameters, unfreeze some or all previously frozen layers (depending on the specific situation and effect), and use a small amount of data from the target subway station to further train the model so that the model can better adapt to the specific characteristics and data distribution of the target subway station.

[0097] S5. Fine-tune the prediction model and perform load forecasting based on the existing data of the target subway station, and use relevant evaluation indicators to evaluate the performance of the model:

[0098] Mean Square Error (MSE):

[0099] in: y pis the true value, is the predicted value, z is the sample size;

[0100] Mean absolute error (MAE):

[0101] Coefficient of determination R 2 :

[0102] in: is the average of the true values.

[0103] By calculating relevant evaluation indicators, we can understand the performance of the model and further optimize or select the model based on the indicators.

[0104] Combine the Transformer model for short-term load forecasting and the LSTM neural network for long-term load forecasting.

[0105] The load forecasting step includes the following sub-steps: using the pre-trained and fine-tuned Transformer model to predict the short-term building load of the target subway station; using the pre-trained and fine-tuned LSTM neural network to predict the long-term building load of the target subway station; combining the prediction results of the Transformer model and the LSTM neural network, comprehensively evaluating and outputting the final prediction value.

[0106] Finally, the prediction results of the two models can be combined through simple averaging, weighted averaging, or more complex ensemble learning methods. Transformer load forecasting is better at short-term forecasting, while LSTM neural network load forecasting is better at long-term load forecasting. Based on this feature, setting relevant weights for the two load forecasting methods can more accurately predict the load. The weights will also change according to the changes in the forecast period. For example, as the expected period increases, the LSTM neural network prediction will be more accurate, and the weight will also increase accordingly. Assume is the Transformer model at time t The predicted value of is the LSTM model at time t The predicted value of w 1 and w 2 ( w 1 + w 2 =1) is the corresponding weight, then the final predicted value can be expressed as:

[0107] Case application:

[0108] Load forecasting based on the data of daily power consumption (kwh) of a subway station throughout the year:

[0109] (1) LSTM neural network load forecasting:

[0110] The training set is set to 90% of the data set, and the test set is set to 10% of the data set. The load forecasting model is built through Matlab and the data is predicted.

[0111] The prediction results are as follows (see Table 1, Table 2 and Figure 4 ):

[0112] (2) Transform load forecasting:

[0113] The training set is set to 90% of the data set, and the test set is set to 10% of the data set. The load forecasting model is built through Matlab and the data is predicted.

[0114] The prediction results are as follows (see Table 3, Table 4 and Figure 5 ):

[0115] (3) LSTM combined with Transform load forecasting

[0116] Combined with the above results, weights are allocated according to the error index. 2 The error index is weighted, LSTM is given a weight of 0.5, and Transform is given a weight of 0.5 to perform a parallel combination prediction. The prediction results are as follows (see Table 5, Table 6 and Figure 6 ):

[0117] By combining the two load forecasting methods, various error evaluation indicators show that the load forecasting accuracy after the combination is significantly better than that of a single method.

[0118] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and technical concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A physical-data dual-driven subway station building load prediction method, characterized in that: The following steps are involved: S1. Use energy consumption simulation software to simulate and generate unit area building load data of subway stations under different meteorological parameters and passenger flow; S2. Decomposing the building load data generated under different meteorological parameters and passenger flow conditions by using variational mode decomposition into several intrinsic mode functions; extracting characteristic parameters from each decomposed intrinsic mode function; S3. Based on the extracted characteristic parameters, the similarity between the target subway station and the loads of multiple subway stations is calculated using the Euclidean distance to obtain a similarity score for the load of each subway station; according to the similarity score calculated using the Euclidean distance, a similarity measurement index is determined, and several subway station load data with the highest similarity scores are selected; S4. Pre-train the initial model using the selected subway station load data, and fine-tune the pre-trained model using a small amount of data from the target subway station to obtain a trained prediction model, so as to improve the accuracy of the prediction model in predicting the building load of the target subway station; S5. Use the trained prediction model to perform load forecasting; Step S2 comprises: Step S21, using variational mode decomposition to decompose the building load data generated under different meteorological parameters and passenger flow conditions into several intrinsic mode functions; the intrinsic mode function is defined as an amplitude-frequency modulation signal : in, is the signal envelope, a non-negative function, and j is an imaginary unit; is the instantaneous phase of the AM / FM signal, ; Step S22, variational mode decomposition estimates the center frequency and bandwidth of each intrinsic mode function by solving a constrained variational problem, and the constrained variational problem is expressed as: in, is the original signal, It is k The intrinsic mode functions, It is k The center frequency of the eigenmode function, represents the convolution operation, is the unit impulse function, Indicates time t The partial derivative of , K is the number of decomposed intrinsic mode functions, is the square of the L2 norm; Step S23, by iterative optimization to minimize the cost function, solve the following optimization problem to update ω k : in, It is represented as the updated cost function after the q+1th iteration; This optimization problem is solved numerically and its solution is expressed as: in, u k ( ω ) yes u k ( t )’s Fourier transform; Step S24: Update the modal function and center frequency until convergence.

2. The physical-data dual-driven subway station building load prediction method according to claim 1 is characterized in that: In step S1, the energy consumption simulation software generates data: meteorological parameters: temperature, humidity and wind speed; Passenger flow: the number of passengers entering and leaving the subway station per unit time; Building load: electricity consumption and air-conditioning energy consumption per unit area.

3. The physical-data dual-driven subway station building load prediction method according to claim 1 is characterized in that: In step S24, the modal function and the center frequency are updated until convergence, including: Update the modal function by solving the following optimization problem : in, is the updated modal function after q+1 iterations, α is a balancing parameter for data fidelity constraints, is the estimated value of the center frequency at the qth iteration; In the fixed mode function In the case of , the center frequency is updated by solving the following optimization problem : .

4. The physical-data dual-driven subway station building load prediction method according to claim 1 is characterized in that: In step S2, the characteristic parameters extracted from each decomposed intrinsic mode function include: (1) Mean: describes the average level of the signal. The calculation formula is: Among them, x n is the nth data point in the IMF component, and N is the total number of data points; (2) Variance: describes the discreteness of the signal. Its calculation formula is: ; (3) Peak value: The maximum value in the signal. Its calculation formula is: ; (4) Kurtosis: describes the sharpness of the signal waveform. Its calculation formula is: ; (5) Effective value: A measure of the effective energy or power of a signal. The calculation formula is: ; (6) Crest factor: The ratio of the peak value to the effective value, used to describe the peak characteristics of the signal. The calculation formula is: ; (7) Pulse factor: The ratio of peak value to average value, used to describe the impulse characteristics of the signal. Its calculation formula is: ; (8) Waveform factor: The ratio of the effective value to the average value, used to describe the waveform characteristics of the signal. The calculation formula is: ; (9) Margin factor: The ratio of peak value to root square amplitude, used to describe the dynamic range of the signal. Its calculation formula is: .

5. The physical-data dual-driven subway station building load prediction method according to claim 1 is characterized in that: Step S3 includes: Step S31: standardize the characteristic data of the target subway station and the subway station load. The formula for standardization is: in, X norm is the standardized feature data, X is the characteristic data containing the target subway station and subway station load, μ is the mean of the feature data, σ is the standard deviation of the feature data; Step S32: Calculate the Euclidean distance of each characteristic data between the target subway station and the subway station load. The calculation formula of the Euclidean distance is: middle: d It is i The Euclidean distance of the features, x i and y i They are points P and Point Q No. i eigenvalues; Step S32: After obtaining the Euclidean distance of each feature, each Euclidean distance is integrated to obtain a score representing the overall similarity: in: S h is the overall similarity score between the load of the hth subway station and the target subway station, z is the number of features; Step S33: Based on the overall similarity score S between the subway station load and the target subway station h , calculate the SMI similarity measure: in, m is the total number of subway station loads, It is h The weight of the subway station load is satisfied ; Step S34: sort the subway station loads according to the SMI similarity measurement index, and select the 100 subway station load data with the highest similarity scores.

6. The physical-data dual-driven subway station building load prediction method according to claim 1 is characterized in that: Step S4 includes the following sub-steps: Step S41: pre-train the initial model using the selected subway station load data, freeze some model layers, and only adjust the last layer to adapt to the target subway station data; Frozen layers: set on the model architecture, the weights of these layers are not updated; Adjustment layer: Assume that the last layer is a fully connected layer, and its output layer weight is W target and the bias is b target , then the output of this layer is expressed as , where Q is the output of the previous layer; Step S42: fine-tune the pre-trained model using a small amount of data from the target subway station and update the model parameters; Step S43: forecast the building load of the target subway station using the fine-tuned model, and evaluate the model performance.

7. The physical-data dual-driven subway station building load prediction method according to claim 6 is characterized in that: In step S4, the prediction model includes a Transformer model for short-term load prediction and an LSTM neural network for long-term load prediction.

8. The physical-data dual-driven subway station building load prediction method according to claim 7 is characterized in that: Combine the prediction results of the Transformer model and the LSTM neural network, comprehensively evaluate and output the final prediction value; Weighted Average: Assumptions is the Transformer model at time t The predicted value of is the LSTM model at time t The final predicted value is expressed as: in, w 1 and w 2 is the corresponding weight, w 1+ w 2=1.

9. The physical-data dual-driven subway station building load prediction method according to claim 7 is characterized in that: Step S5 also includes: using relevant evaluation indicators to evaluate the performance of the model; the evaluation indicators include: Mean Square Error (MSE): in: y p is the true value, is the predicted value, z is the sample size; Mean absolute error (MAE): Coefficient of determination R 2 : in: is the average of the true values.

Images