Dynamic daily load prediction method and system for energy-saving heat supply of public building in winter

Abstract

The invention provides a dynamic daily load prediction method and system for energy-saving heat supply of a public building in winter, and belongs to the field of energy-saving operation of a building energy system.The method comprises the steps that influence factors of heat supply to be predicted are obtained; decomposing each heat supply influence factor to be predicted to obtain a plurality of modes of each influence factor; in the multiple modes of each influence factor, the mode with the highest coefficient of correlation with the heat load is taken, and reconstructed data is formed by the mode and the to-be-predicted heat supply influence factor; and the reconstructed data are input into a pre-trained LSTM neural network model to obtain a daily load prediction result, and the pre-trained LSTM neural network model adopts a sparrow search method to optimize parameters of the LSTM neural network model. According to the method, multivariate variational mode decomposition, the LSTM neural network model and a sparrow search method are adopted, the situation that the thermal load value is too large is avoided, and the prediction efficiency and accuracy are improved.

Description

Technical Field

[0001] The present invention belongs to the field of energy-saving operation of building energy systems, and particularly relates to a dynamic daily load prediction method and system for energy-saving heating in public buildings in winter. Background Art

[0002] In cold regions, due to the extremely large temperature difference between indoor and outdoor in winter, the steady-state heat transfer calculation model is generally used for building heat load calculation. However, the initial outdoor heating design temperature of the energy system is lower than the actual outdoor temperature, which often leads to an overestimated heat load value in the calculation process by traditional methods. The heat load of the air conditioning system in public buildings is affected by various factors, mainly including indoor and outdoor temperature and humidity, the thermal performance of the building envelope, outdoor wind speed, wind direction, and solar radiation. In addition, under the same outdoor meteorological conditions, due to the randomness and uncertainty of factors such as equipment startup, personnel changes, and operation adjustment methods, there will also be significant differences in load prediction. Using the steady-state heat transfer calculation model to calculate the load does not fully consider the coupling effect of various factors and cannot reflect the dynamic response of the building to changes in internal and external parameters. As a result, the winter heat load prediction results are generally too high, and the equipment selection is too large, causing many heat source devices to be idle in actual operation, not only increasing the initial investment cost, but also the equipment operating at low load all year round, with low efficiency of the main engine, water pump, and fan, and being difficult to adjust during the operation stage, resulting in too high indoor temperature during winter heating in the building, and even huge energy waste such as wearing summer clothes in winter and opening windows in winter.

[0003] Currently, commonly used load simulation software is generally based on physical models. However, in actual applications, simplification is often carried out, ignoring the impact of random noise. For example, devices (such as air conditioners, heaters, computers, etc.) do not operate constantly in actual operation but start and stop randomly according to factors such as demand changes, faults, and maintenance. This additional load fluctuation may not be reflected by a simple physical model; changes in meetings, social activities, working hours, etc. will also affect the heat load demand in the building. Many simulation tools assume that the activity level of people is regular or only estimate based on fixed working hours, but this ignores the volatility of the activity level; short-term fluctuations in meteorological conditions will significantly affect the heat load demand of the building. For example, a sudden cold snap occurs, causing the temperature to drop sharply, which may lead to an overestimation or underestimation of the heating demand of the building and unable to meet the actual heat load demand; the dynamic response characteristics of the heating system are also often simplified in actual applications. Many load calculation software assumes that the heat supply of the heating system is instant and constant, but in actual operation, the heating system takes a certain amount of time to respond to changes in indoor temperature. The above-mentioned random noise does not exist in the form of independent data but is hidden in the observable data sequence. The interference components such as equipment start and stop and meteorological mutations can be filtered through signal separation technology, while retaining the core trend characteristics of the physical model. Finally, the decomposed modal components present clear physical meanings and periodic laws.

[0004] In cold regions, energy-saving regulation strategies are adopted for winter heating, which requires accurate prediction of hourly loads to adjust parameters such as outlet water temperature and flow rate. According to statistics, the regulation strategy based on dynamic daily load prediction can save 10% - 20% of energy, or even higher. The changes in monthly and annual loads of buildings are relatively stable, with a larger error tolerance during prediction, and relatively lower requirements for data and models. However, daily load prediction is affected by various short-term dynamic factors, and a large amount of historical data has not been deeply mined. In addition, the daily load has stronger volatility and needs to deal with strong nonlinear and dynamic relationships, which to a certain extent leads to deviations in the selection of energy-consuming equipment and is also not conducive to the operation of the online real-time regulation system. Therefore, it is necessary to continuously explore more accurate and reliable dynamic daily load prediction methods. Summary of the Invention

[0005] In view of the deficiencies of the prior art, this application proposes a dynamic daily load prediction method and system for energy-saving heating in public buildings in winter.

[0006] In a first aspect, this application proposes a dynamic daily load prediction method for energy-saving heating in public buildings in winter, including:

[0007] Obtain the influencing factors of the heating to be predicted;

[0008] Decompose each influencing factor of the heating to be predicted to obtain multiple modes of each influencing factor;

[0009] Among multiple modalities of each influencing factor, select the modality with the highest correlation coefficient with the heat load, and form the reconstructed data with the influencing factors of the heat supply to be predicted.

[0010] Input the reconstructed data into the pre-trained LSTM neural network model to obtain the daily load prediction result. The pre-trained LSTM neural network model is trained using the historical data of the influencing factors of the heat supply, and the sparrow search method is used to optimize the parameters of the LSTM neural network model.

[0011] The training process of the pre-trained LSTM neural network model includes:

[0012] Obtain the historical data of the influencing factors of the heat supply, and use the heat load of the historical building as the prediction label.

[0013] According to the historical data of the influencing factors of the heat supply and the prediction label, calculate the Pearson correlation coefficient, and remove the historical data of the influencing factors of the heat supply with a Pearson correlation coefficient less than the preset threshold to obtain the first historical reconstruction data set.

[0014] Decompose the historical data of each influencing factor in the first historical reconstruction data set to obtain multiple modalities of the historical data of each influencing factor.

[0015] For multiple modalities of the historical data of each influencing factor, calculate the Pearson correlation coefficient, and screen out the corresponding modality with the highest Pearson correlation coefficient.

[0016] Combine the corresponding modality with the highest Pearson correlation coefficient with the historical data of the influencing factors of the heat supply to form the second historical reconstruction data set.

[0017] Train the LSTM neural network model using the second historical reconstruction data set to obtain the pre-trained LSTM neural network model.

[0018] The optimization of the parameters of the LSTM neural network model using the sparrow search method includes:

[0019] Normalize the second historical reconstruction data set.

[0020] Take the maximum number of training times, the number of hidden units, and the initial learning rate in the pre-trained LSTM neural network model as the training objectives of the sparrow search method. Based on the normalized second historical reconstruction data set, use the sparrow search method to iteratively train the maximum number of training times, the number of hidden units, and the initial learning rate in the trained LSTM neural network model to obtain the optimal maximum number of training times, the optimal number of hidden units, and the optimal initial learning rate.

[0021] Construct an optimized pre-trained LSTM neural network model with the optimal maximum number of training times, the optimal number of hidden units, and the optimal initial learning rate.

[0022] The Pearson correlation coefficient is calculated as follows:

[0023]

[0024] where γ X,Y is the Pearson correlation coefficient between the historical data of the influencing factors of heat supply and the prediction labels, m is the number of historical data of the influencing factors of heat supply, X r is the r-th historical data of the influencing factors of heat supply, Y r is the r-th prediction label, that is, the heat load of the historical building, is the average value of the historical data of the influencing factors of heat supply, is the average value of the prediction labels, γ X,Y ranges from [-1, 1]. When γ X,Y is close to 1 or -1, it indicates a higher correlation between the two variables X and Y. When γ X,Y is closer to 0, it indicates a lower correlation between the two variables X and Y.

[0025] Decompose each piece of historical data of the influencing factors of heat supply in the first historical reconstruction dataset to obtain multiple modes of the historical data of each influencing factor, including:

[0026] Use each piece of historical data of the influencing factors of heat supply in the first historical reconstruction dataset as the input data of the data channel, and use multiple modes of the historical data of each influencing factor as the output data;

[0027] Obtain a multivariate modulation oscillation set according to the relationship between the input data and the output data;

[0028] Perform Hilbert transform on each element in the multivariate modulation oscillation set to obtain a vector analytic signal, and the vector analytic signal consists of the central frequency of the vector analytic signal and the frequency domain representation of the k-th mode in the c-th channel;

[0029] Initialize the Lagrange multiplier, the central frequency of the vector analytic signal, and the frequency domain representation of the k-th mode in the c-th channel;

[0030] Update the frequency domain representation of the k-th mode in the c-th channel using the Lagrange multiplier and the penalty factor;

[0031] Update the central frequency using the frequency domain representation of the k-th mode in the c-th channel;

[0032] Update the Lagrange multiplier using the noise tolerance;

[0033] When the preset convergence condition is reached, the update of the Lagrange multiplier, the center frequency, and the frequency-domain representation of the k-th mode in the c-th channel is terminated, and the frequency-domain representation of the k-th mode in the c-th channel after the (n + 1)-th iteration is output. The frequency-domain representation of the k-th mode in the c-th channel after the (n + 1)-th iteration corresponds to multiple modes of the historical data of each updated influencing factor.

[0034] The frequency-domain representation of the k-th mode in the c-th channel is updated by using the Lagrange multiplier and the penalty factor, and the calculation formula is as follows:

[0035]

[0036] where n is the maximum number of iterations, α is the penalty factor, λ c (ω) is the Lagrange multiplier, ω is the frequency variable, ω k is the center frequency of the k-th mode, x c (ω) is the frequency-domain representation of the input data of the c-th channel, u k,c (ω) is the frequency-domain representation of the k-th mode in the c-th channel, is the frequency-domain representation of the k-th mode in the c-th channel after the (n + 1)-th iteration.

[0037] The center frequency is updated by using the frequency-domain representation of the k-th mode in the c-th channel, and the calculation formula is as follows:

[0038]

[0039] where c is the number of data channels, u k,c (ω) is the frequency-domain representation of the k-th mode in the c-th channel, is the center frequency of the k-th mode after the (n + 1)-th iteration.

[0040] The Lagrange multiplier is updated by using the noise tolerance, and the calculation formula is as follows:

[0041]

[0042] where τ is the noise tolerance, is the frequency-domain representation of the k-th mode in the c-th channel after the (n + 1)-th iteration, is the Lagrange multiplier of the c-th channel at the n-th iteration.

[0043] In a second aspect, the present application provides a dynamic daily load prediction system for winter energy-saving heating of public buildings, including:

[0044] A data acquisition module, configured to acquire the influencing factors of the heating to be predicted;

[0045] A modal decomposition module, which is used to decompose each influencing factor of heat supply to be predicted to obtain multiple modes of each influencing factor;

[0046] A data reconstruction module, which is used to select the mode with the highest correlation coefficient with the heat load from multiple modes of each influencing factor, and form reconstructed data with the influencing factor of heat supply to be predicted;

[0047] A prediction result output module, which is used to input the reconstructed data into a pre-trained LSTM neural network model to obtain a daily load prediction result. The pre-trained LSTM neural network model is trained using historical data of the influencing factors of heat supply, and the sparrow search method is used to optimize the parameters of the LSTM neural network model.

[0048] In a third aspect, the present application proposes an electronic device, including: one or more processors, and a memory. The memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the dynamic daily load prediction method for energy-saving heating in winter of public buildings.

[0049] In a fourth aspect, the present application proposes a computer-readable storage medium, which stores executable instructions. When the instructions are executed, the processor executes the dynamic daily load prediction method for energy-saving heating in winter of public buildings.

[0050] In a fifth aspect, the present application proposes a computer program product, including a computer program or instructions. When the computer program or instructions are executed by a processor, the dynamic daily load prediction method for energy-saving heating in winter of public buildings is implemented.

[0051] Beneficial effects:

[0052] The present application proposes a dynamic daily load prediction method and system for energy-saving heating in winter of public buildings. Aiming at the situation that the heat load value is often too large in the existing calculation process, the LSTM neural network model adopted by the present application can effectively mine the hidden coupling relationship between input parameters and extract the time correlation of features, avoiding the situation of too large heat load value. The present application also adopts a sparrow search parameter optimization mechanism, avoiding the blindness and uncertainty of manual setting, and further improving the efficiency and accuracy of the LSTM neural network model. Aiming at the problem that the random noise in the existing model does not exist in the form of independent data, but is hidden in the observable data sequence, the present application adopts multivariate variational mode decomposition to denoise the data. This step not only enhances the model's ability to process complex data, but also provides a more reliable data basis for subsequent prediction. Description of the Drawings

[0053] Figure 1 Flowchart of a dynamic daily load prediction method for winter energy-saving heating of public buildings in an embodiment of the present application;

[0054] Figure 2 Schematic diagram of the flow of a dynamic daily load prediction method for winter energy-saving heating of public buildings in an embodiment of the present application;

[0055] Figure 3 Schematic diagram of the training process of the LSTM neural network model in an embodiment of the present application;

[0056] Figure 4 Schematic diagram of the sparrow search optimization method in an embodiment of the present application;

[0057] Figure 5 Result diagram of outdoor temperature modal decomposition when K = 8 in an embodiment of the present application;

[0058] Figure 6 Result diagram of outdoor relative humidity modal decomposition when K = 8 in an embodiment of the present application;

[0059] Figure 7 Result diagram of solar radiation illuminance modal decomposition when K = 8 in an embodiment of the present application;

[0060] Figure 8 Result diagram of the in-room occupancy rate modal decomposition when K = 8 in an embodiment of the present application;

[0061] Figure 9 Result diagram of building load modal decomposition when K = 8 in an embodiment of the present application;

[0062] Figure 10 Prediction error of building heat load in different decomposition layers in an embodiment of the present application, where (a) is a schematic diagram of RMSE error, (b) is a schematic diagram of MAE error, (c) is a schematic diagram of R 2 Error schematic diagram, and (d) is a schematic diagram of VAR error;

[0063] Figure 11 Comparison diagram of model prediction results on December 15th in an embodiment of the present application. Detailed implementation manners

[0064] The following further describes in detail the specific implementation manners of the present application in conjunction with the drawings and embodiments.

[0065] In the calculation process of traditional methods, the heat load value often becomes too large. Traditional methods include manual calculation and software calculation. The reason for the large load value using this method is steady-state calculation. Steady-state calculation generally calculates the heat load according to the lowest outdoor temperature through formulas according to regulations, which will be on the high side. While the present application uses a neural network model, which does not require calculation, but is inferred through training with historical data by exploring the rules between data.

[0066] For the traditional model constructed based on the steady-state heat transfer differential equation commonly used in heat load calculation in cold regions, too many actual conditions are simplified. Coupled with the large range of ambient temperature changes in cold regions, the load simulation results tend to be on the high side. An innovative short-term daily building load prediction method based on MVMD-SSA-LSTM is proposed.

[0067] The randomness of the factors affecting building heat load leads to the problem of noise interference in the time series data of building heat load. Therefore, the Multivariate Variational Mode Decomposition (MVMD) algorithm is introduced to denoise the data. Its core lies in its ability to handle the decomposition of multivariate signals, which can be decomposed in both the time and frequency dimensions. By using the co-variation between different signals, it can more accurately extract and separate modal components. Moreover, MVMD has the characteristic of adaptive frequency decomposition, and it can automatically assign a suitable frequency band to each mode during calculation, greatly reducing the phenomenon of modal aliasing. After decomposition, the mode with the highest correlation with the output data is extracted as a supplementary feature for prediction, thus effectively improving the accuracy and stability of load prediction. This processing flow not only enhances the model's ability to handle complex data but also provides a more reliable data basis for subsequent prediction.

[0068] The factors affecting building heat load do not act alone but are coupled and jointly influential. For example, during working hours in public buildings, there are a large number of people, and internal loads such as equipment and lighting are significant. These factors will reduce the heating demand. However, if the external climate suddenly gets cold, the heat generated by relying on human activities and equipment is not enough to maintain a comfortable temperature, resulting in an increase in heat load. In addition, as solar radiation increases and temperature decreases, there are cases where the impact of temperature on the load is greater than that of solar radiation on the load. At this time, the building requires more heat load to maintain a comfortable indoor environment. In winter, the internal temperature change of the building lags behind the external temperature change. When the external temperature drops, the internal temperature of the building will not drop immediately, and this lag effect will also affect the calculation of heat load. LSTM can effectively mine the hidden coupling relationships between input parameters and extract the time correlation of features. The core advantage of this algorithm lies in its Constant Error Carousel (CEC) and gating mechanism. It can maintain a stable transfer of information when processing long time series by forcing a constant error flow inside the network, avoiding learning difficulties caused by gradient decay or explosion. In addition, it can flexibly control which information should be retained and which should be forgotten. Compared with the traditional physical model calculation method, it can accurately reflect the fluctuation characteristics of the load, save computing resources, and have better prediction effects.

[0069] During the model construction process, a parameter optimization mechanism of the Sparrow Search Algorithm (SSA) was designed to automatically search for and determine the optimal settings of the three main parameters of the Long Short-Term Memory Neural Network (LSTM). This avoided the blindness and uncertainty of manual setting and further improved the efficiency and accuracy of the prediction model. SSA not only includes the exploration behavior of individuals but also a group cooperation mechanism. This dynamically adaptive search method can find more suitable solutions in various complex scenarios. It does not rely on fixed global search strategies or local search strategies and can adaptively adjust the balance between the two during the search process. Especially in high-dimensional and multimodal optimization problems, SSA can adaptively balance global search and local search. Finally, through training and verification with input data, the model successfully achieved accurate prediction of the short-term daily load of buildings, providing accurate data for real-time energy optimization.

[0070] To achieve efficient and accurate prediction of building daily load and provide data support for energy management and control of large public buildings, the following technical solutions are proposed in this study. By mainly using noise reduction technology, optimization technology, and deep learning technology, the main influencing factors of load prediction are screened. After refined modal decomposition to achieve data noise reduction, supplementary features for prediction are screened, and the parameters of the neural network are optimized to establish a combined MVMD-SSA-LSTM building dynamic daily load prediction model.

[0071] Example 1:

[0072] This example proposes a dynamic daily load prediction method for winter energy-saving heating of public buildings, as Figure 1 、 Figure 2 shown, including:

[0073] Step S1: Obtain the influencing factors of the heating to be predicted;

[0074] In this example, the influencing factors of the heating to be predicted include six main influencing factors: outdoor temperature, outdoor humidity, solar radiation illuminance, occupancy rate, wind speed, and wind direction, but are not limited to the above influencing factors.

[0075] Step S2: Decompose each influencing factor of the heating to be predicted to obtain multiple modes of each influencing factor;

[0076] In this example, the MVMD algorithm is used for decomposition. The parameters of the penalty factor, noise tolerance, initialization mode, convergence value, and number of modes are set, and the decomposition results when the number of modes is 2-10 are saved respectively to obtain multiple modes of each influencing factor.

[0077] Step S3: Among multiple modalities of each influencing factor, select the modality with the highest correlation coefficient with the heat load, and form the reconstructed data with the influencing factors of the heat supply to be predicted.

[0078] In this embodiment, the correlation coefficient with the heat load is calculated using the Pearson correlation coefficient, and the one with the highest value is selected among all the Pearson correlation coefficients. The modality with the highest Pearson correlation coefficient and the influencing factors of the heat supply to be predicted together constitute the reconstructed data.

[0079] Step S4: Input the reconstructed data into the pre-trained LSTM neural network model to obtain the daily load prediction result. The pre-trained LSTM neural network model is trained using the historical data of the influencing factors of the heat supply, and the sparrow search method is used to optimize the parameters of the LSTM neural network model.

[0080] In this embodiment, the reconstructed data is input into the pre-trained LSTM neural network model to obtain the final daily load prediction result. Among them, various parameters in the pre-trained LSTM neural network model are obtained after being optimized by the sparrow search method. The various parameters at least include: the maximum number of training times, the number of hidden units, and the initial learning rate.

[0081] In step S4, the pre-trained LSTM neural network model, as Figure 3 shown, the training process includes:

[0082] Step S4.1: Obtain the historical data of the influencing factors of the heat supply, and use the heat load of the historical building as the prediction label.

[0083] This implementation case is a large public building in Shenyang: The effective heating area of the building in winter is about 80,000 square meters. The air-conditioning system is turned on from 7:00 to 19:00. A building energy monitoring platform has been installed, and the recording interval of each influencing factor and the building heat load data is 1 hour.

[0084] Select the historical building heat load data and its related influencing factor data during the heating seasons (November - March of the following year) in 2021 and 2022, with a total of 3,600 data samples. The influencing factors include six main influencing factors: outdoor temperature, outdoor humidity, solar radiation irradiance, occupancy rate, wind speed, and wind direction. Save the data in the form of a table.

[0085] Step S4.2: Calculate the Pearson correlation coefficient according to the historical data of the influencing factors of the heat supply and the prediction label, and remove the historical data of the influencing factors of the heat supply with a Pearson correlation coefficient less than the preset threshold to obtain the first historical reconstruction data set.

[0086] For the obtained historical building heat load data and its related influencing factor data, a decomposition result diagram is drawn. At the same time, the Pearson correlation coefficient between each influencing factor and the building heat load is calculated, and the weakly correlated influencing factors are screened out. The remaining influencing factors are used as the original influencing factors for subsequent prediction, and a reconstructed dataset is formed with the building heat load. The weakly correlated influencing factors will not participate in the subsequent steps. Among them, the preset threshold can be set according to the actual situation.

[0087] Specifically, the Pearson correlation coefficient can be calculated by Equation (1), and its strength can be represented by Table 1:

[0088]

[0089] Among them, γ X,Y is the Pearson correlation coefficient between the historical data of the influencing factors for heating and the prediction labels, m is the number of historical data of the influencing factors for heating, X r is the r-th historical data of the influencing factors for heating, Y r is the r-th prediction label, that is, the heat load of the historical building. is the average value of the historical data of the influencing factors for heating. is the average value of the prediction labels, γ X,Y ranges from [-1, 1]. When γ X,Y is close to 1 or -1, it indicates that the correlation between the two variables X and Y is higher. When γ X,Y is closer to 0, it indicates that the correlation between the two variables X and Y is lower.

[0090] Calculate the Pearson correlation coefficients between the data of the above six influencing factors and the building heat load data respectively according to Equation (1), and screen out the influencing factors with weak correlation and below according to the specified values in Table 1. Table 2 shows the calculation results of this step for the case. Implement the above steps to screen out the two influencing factors of wind speed and wind direction. At this time, the influencing factors of the building heat load are outdoor temperature, outdoor humidity, solar radiation illuminance, and occupancy rate.

[0091] Table 1 Pearson correlation coefficient strength table

[0092] Value of |γ| Degree of correlation 0.1 ≤ |γ| < 0.3 Weak correlation 0.3 ≤ |γ| < 0.5 Moderate correlation 0.5 ≤ |γ| < 0.7 Relatively strong correlation 0.7 ≤ |γ| < 0.9 Strong correlation 0.9 ≤ |γ| ≤ 1 Very strong correlation

[0093] Table 2 Pearson correlation coefficients between each influencing factor and the building heat load in the implementation case

[0094]

[0095]

[0096] Thus, the outdoor temperature, outdoor humidity, solar radiation illuminance, and occupancy rate form the first historical reconstruction dataset with the building heat load.

[0097] Step S4.3: Decompose the historical data of each heat supply influencing factor in the first historical reconstruction dataset to obtain multiple modes of the historical data of each influencing factor.

[0098] In this embodiment, the historical data of each heat supply influencing factor in the first historical reconstruction dataset is decomposed using the MVMD algorithm. Parameters such as the penalty factor, noise tolerance, initialization mode, convergence value, and number of modes are set, and the decomposition results when the number of modes is 2 - 10 are saved respectively. After decomposition, the Pearson correlation coefficient between each mode (IFM) and the building heat load is calculated, and the mode with the highest factor coefficient for each factor is used as a supplementary feature for load prediction. After extracting the features, the dataset is reconstructed again, and the decomposition result of the building heat load will be defined as the prediction output.

[0099] Specifically, when the number of decomposition modes is small, there may be information loss, and important frequency components and signal features may not be fully captured; when the number of modes is large, noise and random fluctuations in the data will be captured, thus reducing the generalization ability of the model and increasing the computational complexity at the same time. Therefore, to avoid the phenomenon of being too large or too small, the number of modes K can be set to 2 - 8, but it is not limited to this range. In addition, the specific decomposition method of the MVMD decomposition algorithm follows the following steps:

[0100] Step S4.3.1: Use the historical data of each heat supply influencing factor in the first historical reconstruction dataset as the input data of the data channel, and use multiple modes of the historical data of each influencing factor as the output data.

[0101] In this embodiment, the input data containing C data channels (i.e., C heat supply influencing factors) is expressed as x(t) = [x1(t), x2(t)...x C (t)], and is decomposed into K preset multivariate modulated oscillations u k (t).

[0102]

[0103] Among them, u k (t) = [u1(t), u2(t),...u C (t)].

[0104] Step S4.3.2: Obtain a set of multivariate modulated oscillations according to the relationship between the input data and the output data.

[0105] Step S4.3.3: Perform Hilbert transform on each element in the set of multivariate modulated oscillations to obtain a vector analytic signal, and the vector analytic signal is composed of the central frequency of the vector analytic signal and the frequency domain representation of the kth mode in the cth channel.

[0106] Obtain the multivariate modulation oscillation set according to Equation (2). Use the principle of Hilbert transform to transform each component of u k (t) to obtain the vector analytic signal For multivariate modulation oscillation, this model assumes that there is a common component ω in all data channels k , in order to make the frequency components of each mode function concentrated around the predetermined center frequency ω k nearby, it is necessary to multiply by a complex exponential according to the principle of Fourier transform for harmonic mixing, and calculate the bandwidth of the multivariate oscillation according to the obtained matrix through the Frobenius norm, expressed by Equation (3):

[0107]

[0108] where is the vector analytic signal corresponding to the c-th channel and the k-th mode

[0109] It can be understood that the frequency-domain representations of these modes decompose the multi-dimensional signal through iterative optimization. Each mode can be regarded as a band-limited component of the original vector analytic signal. The center frequencies of these modes are adjusted around their respective center frequencies and jointly form the complete vector analytic signal. That is, one u k,c (ω) corresponds to one IMF in the figure

[0110] Step S4.3.4: Initialize the Lagrange multiplier, the center frequency of the vector analytic signal, and the frequency-domain representation of the k-th mode in the c-th channel

[0111] In this embodiment, during calculation, it is necessary to initialize three functions to be updated, namely {u k,c}, {ω k}, λ c . Set the initialization mode init to: all initial frequencies are evenly distributed, that is, initialize according to the method of ω = (0.5 / K) × ((1:K)-1), and the rest of the variables are all initialized to 0

[0112] Step S4.3.5: Update the frequency-domain representation of the k-th mode in the c-th channel using the Lagrange multiplier and the penalty factor

[0113] In this embodiment, mode update and center frequency update are its core calculation steps, mainly used to accurately extract the modal characteristics in the signal, reduce the modal aliasing problem of the decomposition result, and improve the decomposition accuracy and robustness of the algorithm. Mode update is the process of updating the mode function in each iteration, as shown in Equation (4):

[0114]

[0115] Among them, n is the maximum number of iterations; α is the penalty factor, which is used to control the smoothness in modal decomposition. The larger the value, the smoother the mode; λ c (ω) is the Lagrange multiplier, which represents the sensitivity of the constraint to the objective function. ω is the frequency variable, and ω k is the central frequency of the k-th mode, and x c (ω) is the representation of the input data of the c-th channel in the frequency domain, and u k,c (ω) is the frequency-domain representation of the k-th mode in the c-th channel. is the frequency-domain representation of the k-th mode in the c-th channel after the (n + 1)-th iteration.

[0116] Step S4.3.6: Update the central frequency using the frequency-domain representation of the k-th mode in the c-th channel;

[0117] Updating the central frequency can adjust the frequency characteristics of the modal function, making it better adapt to the frequency-domain characteristics of the original signal and improving the decomposition accuracy. The central frequency update function is shown in Equation (5):

[0118]

[0119] Among them, c is the number of data channels, and u k,c (ω) is the frequency-domain representation of the k-th mode in the c-th channel. is the central frequency of the k-th mode in the (n + 1)-th iteration.

[0120] Step S4.3.7: Update the Lagrange multiplier using the noise tolerance;

[0121] The Lagrange multiplier update function is shown in Equation (6):

[0122]

[0123] Among them, τ is the noise tolerance, which represents the noise tolerance and is used for the update of the Lagrange multiplier λ c . is the frequency-domain representation of the k-th mode in the c-th channel in the (n + 1)-th iteration. is the Lagrange multiplier of the c-th channel in the n-th iteration.

[0124] Step S4.3.8: When the preset convergence condition is reached, end the update of the Lagrange multiplier, the central frequency, and the frequency-domain representation of the k-th mode in the c-th channel, and output the frequency-domain representation of the k-th mode in the c-th channel after the (n + 1)-th iteration. The frequency-domain representation of the k-th mode in the c-th channel after the (n + 1)-th iteration corresponds to multiple modes of the historical data of each updated influencing factor.

[0125] In this embodiment, the calculation is continuously updated. Within the maximum number of calculations, when the result is less than the convergence value ∈, the calculation ends. The convergence formula is shown in (7):

[0126]

[0127] Thus, the decomposition results u of each factor can be obtained k,c , that is, (IMFs). The decomposition results of the first factor are respectively: u 1,1 , u 2,1 , u 3,1 ······u K,1 , corresponding to IMF1, IMF2, IMF3······IMF K of the decomposition result of the first factor. The decomposition results of the second factor are respectively: u 1,2 , u 2,2 , u 3,2 ······u K,2 , corresponding to IMF1, IMF2, IMF3······IMF K of the decomposition result of the second factor, and so on.

[0128] It should be mainly noted that: the reconstructed data is input into the pre-trained LSTM neural network model to obtain the daily load prediction result. Among them, the daily load prediction result is obtained by adding the predicted values obtained each time. The reconstructed data includes K modes and residual data. When predicting, each mode and residual are used as the output of the prediction for one prediction, and the final prediction result is obtained by adding the predicted values obtained each time.

[0129] Step S4.4: Calculate the Pearson correlation coefficient for multiple modes of the historical data of each influencing factor, and select the corresponding mode with the highest Pearson correlation coefficient;

[0130] Step S4.5: Combine the corresponding mode with the highest Pearson correlation coefficient and the historical data of the influencing factors of heating to form a second historical reconstruction data set;

[0131] In this embodiment, the maximum number of calculations n value and the penalty factor value α in formula (3) are set; the noise tolerance in formula (5) is set; the convergence value ∈ in formula (6) is set; the number of modes K is set; the specific settings are shown in Table 3, and the iterative calculation of the above step (3) is performed to perform MVMD decomposition on the reconstruction data set.

[0132] Table 3 Parameter settings of the MVMD algorithm in the implementation case

[0133]

[0134] Save the decomposition results when the number of modes K is from 2 to 10, including the K modes of each factor and their residual curves. This step is demonstrated with K = 8 as an example, see Appendix Figure 5 、 Figure 6 、 Figure 7 、 Figure 8 、 Figure 9 , after decomposition, calculate the Pearson correlation coefficient between each mode (IMF1 - IMF8) of each influencing factor and the building heat load according to Equation (1). Select the mode with the highest correlation coefficient for each factor as the supplementary feature for load prediction. According to the calculation results shown in Table 4, when K = 8, extract the 1st sub-mode of the outdoor temperature factor, the 1st sub-mode of the outdoor humidity factor, the 2nd sub-mode of the solar radiation irradiance factor, and the 3rd sub-mode of the occupancy rate factor as the supplementary features for prediction. Then reconstruct the dataset again.

[0135] Table 4 Pearson correlation coefficients between each mode of each factor and the building heat load when K = 8

[0136] Mode Outdoor temperature Outdoor humidity Solar irradiance Occupancy rate <![CDATA[1(IMF1)]]> 0.8316 0.6630 0.4168 0.01224 <![CDATA[2(IMF2)]]> 0.3437 0.6300 0.7913 0.4069 <![CDATA[3(IMF3)]]> 0.4919 0.5122 0.2087 0.7597 <![CDATA[4(IMF4)]]> 0.1136 0.3428 0.2164 0.7596 <![CDATA[5(IMF5)]]> 0.0751 0.2348 0.1863 0.2805 <![CDATA[6(IMF6)]]> 0.0581 0.1736 0.1750 0.3079 <![CDATA[7(IMF7)]]> 0.0396 0.0670 0.1742 0.2490 <![CDATA[8(IMF8)]]> 0.0504 0.0554 0.188 0.1412

[0137] Step S4.6: Train the LSTM neural network model using the second historical reconstruction dataset to obtain a pre-trained LSTM neural network model.

[0138] In this embodiment, the mode corresponding to the highest Pearson correlation coefficient and the historical data of the influencing factors of heating are combined to form the second historical reconstruction dataset, and the LSTM neural network model is trained using the second historical reconstruction dataset to obtain a pre-trained LSTM neural network model.

[0139] Set the structural parameters of the LSTM neural network, including the network structure framework, the number of hidden layers, and the number of hidden units; set the prediction operation parameters, including the maximum number of training times, the batch size of iterative samples, the gradient threshold, the initial learning rate, the learning rate decay period, the learning rate decay factor, the regularization parameter, and the training environment.

[0140] Specifically, the network structure framework of the LSTM includes an input layer, an LSTM layer with one hidden unit (processing input sequence data and memorizing long-term and short-term dependency information), a Relu activation layer (introducing non-linearity to enhance the expression ability of the network), a fully connected layer, and a regression layer (calculating the loss for guiding the training and optimization of the network). The Adam optimization algorithm is used for the training option. This network mainly consists of three parts, and its calculation and update steps follow the following formulas:

[0141] Forgetting gate: Responsible for determining how much of the previous cell state is retained to the current cell state, that is, determining what information to discard from the cell state. The calculation formula is shown in Equation (8):

[0142] f t = σ(W f · [h t-1 , x t + b f ) (8)

[0143] Among them, x t is the input vector at time t; C t is the memory at time t, which is essentially the global information from time 0 to t. Using the input x t and the hidden state information h t-1 from the previous time to generate three gating information and a candidate memory: forget gate f t , input gate i t , output gate o t , candidate memory σ represents the sigmoid activation function, W f is the weight matrix of the forget gate, b f is the bias of the current forget gate. This gate reads h t-1 and x t , and then outputs a number f t between 0 and 1 after passing through the sigmoid layer. f t-1 is multiplied element-wise with each number in the cell state C t . A value of 0 for f t means complete discard, and 1 means complete retention.

[0144] Input gate: Determines how much information at the current time should be retained and can be input into the cell state at the current time. It mainly consists of two parts. The first part is the sigmoid (σ) layer, which determines the value to be updated. The second part is the tanh layer, which can update the information to be updated into the cell state. The tanh layer creates a new vector of cell state values (candidate memory), which will be added to the state, as shown in equations (9) and (10):

[0145] i i = σ(W t-1 · [h t , x i ) (9)

[0146]

[0147] Among them, W i , W C are the weight matrices of the input gate and the memory module respectively; b i , b C are the biases between the current input gate and the memory module respectively; C t-1 is updated to C t, multiply the old state by f t , discard the information that needs to be discarded, and then add i t *C t , which completes the update of the cell state, C t The update is shown in Equation (11):

[0148]

[0149] Output gate: A sigmoid layer is used to determine which part of the cell state will be output. The cell state is processed through tanh to obtain a value between -1 and 1, and it is multiplied by the output of the sigmoid gate. Finally, only the determined output part is output. The calculation formulas are shown in Equations (12) and (13):

[0150] o t = σ(W o [h t-1 , x t ) + b o ) (12)

[0151] h t = o t *tanh(C t ) (13)

[0152] Among them, W o is the weight matrix of the output gate, and b o is the bias of the current forget gate.

[0153] Under the Adam algorithm, the update of all weights (W f , W i , W C , W o ) follows Equation (14); the update of all biases (b f , b i , b C , b o ) follows Equation (15):

[0154]

[0155] Among them, i represents the i-th iteration; W i-1 and b i-1 represent the current weights and biases, and W i and b i represent the weights and biases after iterative update; ∈' is the minimum constant to prevent the denominator from being zero. η' is the current learning rate, and the update formula is shown in Equation (16); and are the corrected first-order momentum and second-order momentum, and the correction formulas are shown in Equations (17) and (18):

[0156] η' = η · decay factor (16)

[0157]

[0158] where m i and v i are the first-order momentum and the second-order momentum respectively. The update formulas are shown in Equations (19) and (20); β1 and β2 are the momentum decay rates (β1 is taken as 0.9 and β2 is taken as 0.999); η is the learning rate of the previous learning rate decay period, decay factor is the learning rate decay factor, and the learning rate decay period is the period for adjusting the learning rate each time (once every 60 epochs), and the initial learning rate is set to 10 -7 .

[0159] m i = β1 · m i-1 + (1 - β1) · g i (19)

[0160]

[0161] where g i is the gradient at the i-th iteration, representing the derivative of the loss function with respect to the weights. The gradient threshold is set to 1 to limit the gradient magnitude, m i-1 is the first-order momentum at the (i - 1)-th iteration, and v i-1 is the second-order momentum at the (i - 1)-th iteration. The calculation formula is shown in Equation (21):

[0162]

[0163] where L is the loss function, and the calculation formula of the loss function is shown in Equation (22):

[0164] L total = L data + λ∑W 2 (22)

[0165] where L total is the total loss, including the data loss and the regularization loss, and is the optimization objective function; L data is the data loss, measuring the error between the model prediction value and the true value; λ is the regularization parameter, controlling the influence degree of the regularization term on the total loss to prevent overfitting; ∑W 2 is the sum of squared weights (i.e., the sum of the squares of all weights, and all weights include: W f , W i , W C , W o ), used to limit the magnitude of the weights.

[0166] The error metrics supporting the prediction model include at least four evaluation metrics, namely Root Mean Square Error (RMSE), Coefficient of Determination (R-squared, R 2 ), Mean Absolute Error (MAE), and Variance (VAR).

[0167] Specifically, the calculation of the four error evaluation metrics follows formulas (23) - (26):

[0168]

[0169] In this embodiment, the re - reconstructed dataset (i.e., the second historical re - constructed dataset) under different numbers of modes is defined for input and output, the training set and the test set are divided, the dataset is input into the LSTM neural network, the model is continuously adjusted, the prediction results under different numbers of modes are obtained, and according to the error metrics described in step 5 above, the number of modes with the smallest prediction error under each metric is comprehensively selected to determine the number of modes for subsequent prediction. Specifically, the training set and the test set are divided in a ratio of 7:3, but this ratio is not limited to this value.

[0170] In this embodiment, the second historical re - constructed dataset is input into the LSTM main prediction network. The number of training times, the sample batch size for each iteration are set to determine the total number of iterations; the learning rate decay factor in formula (16) is set, and the corresponding initial learning rate and learning rate decay period are set; the gradient threshold corresponding to the gradient in formula (21) is set; the regularization parameter in formula (22) is set. Table 5 shows the setting results of the LSTM neural network parameters.

[0171] Table 5 Parameter Settings of the LSTM Neural Network Algorithm in the Embodiment

[0172]

[0173] The prediction results are expressed in terms of the error metrics RMSE, MAE, R 2 and VAR, and their calculation follows formulas (23) - (26).

[0174] Columns 1 - 8 of the re - reconstructed dataset are defined as the original influencing factors and each supplementary feature data as the prediction input, and the 9th column is defined as the actual building heat load value. Each mode after the decomposition of the building heat load is used as the output of the load prediction in turn. The training set and the test set are divided in a ratio of 7:3. The dataset is input into the LSTM neural network, and predictions are made at different numbers of modes from 2 to 8, and the prediction results at different modes in this case can be obtained from the appendix Figure 10It shows that, among them, (a) is the schematic diagram of RMSE error, (b) is the schematic diagram of MAE error, (c) is the R 2 schematic diagram of error, and (d) is the schematic diagram of VAR error; in this case, when the number of modes is 8, the four error values perform the best, so the number of modes for subsequent predictions is set to 8.

[0175] In step S4, it also includes using the sparrow search method to optimize the parameters of the LSTM neural network model, as Figure 4 shown, the specific process is as follows:

[0176] Step S4.7: Normalize the second historical reconstruction dataset;

[0177] Step S4.8: Take the maximum number of training times, the number of hidden units, and the initial learning rate in the pre-trained LSTM neural network model as the training objectives of the sparrow search method. Based on the normalized second historical reconstruction dataset, use the sparrow search method to iteratively train the maximum number of training times, the number of hidden units, and the initial learning rate in the trained LSTM neural network model to obtain the optimal maximum number of training times, the optimal number of hidden units, and the optimal initial learning rate;

[0178] In this embodiment, the initial number of iterations is 50, the population is 30, the proportion of discoverers is 0.7, the proportion of sparrows aware of danger is 0.2, the warning value is 0.6, and the dimension of the optimization parameter is 3 (the objectives include the maximum number of training times, the number of hidden units, and the initial learning rate of the LSTM neural network);

[0179] Calculate the fitness value according to the objective function. The fitness values of all sparrows are represented by Equation (27), and the fitness is sorted to find the current best fitness individual and the worst fitness individual.

[0180]

[0181] Among them, n' is the number of sparrows; d is the dimension of the problem variables to be optimized, F X is the fitness matrix of the sparrow population, and x n',d is the position coordinate of the n'th sparrow in the d'th dimension.

[0182] Update the position of the discoverer, which is represented by Equation (28):

[0183]

[0184] where, t' is the current iteration number; j' = 1, 2,... d. iter max is the maximum number of iterations, which is a constant; To represent the position information in the j'-th dimension of the i'-th sparrow; α' ∈ (0, 1] is a random number; R2 ∈ [0, 1] represents the warning value; S T ∈ [0.6, 1] represents the safety value; Q is a random number following a normal distribution; L represents a 1×d matrix, and all elements in this matrix are 1.

[0185] The update of the follower's position can be described by Equation (29):

[0186]

[0187] Among them, exp represents the optimal position where the discoverer is currently located; represents the current globally worst position; is the j'-th dimension position of the i'-th follower at time t'+1, is the position of the p-th follower at time t'+1, is the j'-th dimension position of the i'-th follower at time t'. A is a 1×d matrix, and each element of this matrix is randomly assigned 1 or -1, A + = A T (AA T ) -1 ; when i' > n' / 2, it means that the i'-th joiner has a lower fitness value and does not obtain food, and is in a hungry state. Therefore, it needs to find food elsewhere.

[0188] The random update of the positions of some sparrows (guards) is represented by Equation (30);

[0189]

[0190] Among them, represents the current globally optimal position; β is the step size control parameter, following a normal distribution with a mean of 0 and a variance of 1; D ∈ [-1, 1], is a random number; f i' is the fitness value of the current sparrow individual; f w is the current globally worst fitness value; f g is the current globally best fitness value.

[0191] After each optimization is completed, the parameters will be transmitted to the LSTM, and the LSTM network will be trained according to the updated parameter combination, and the fitness will be calculated based on the results of the LSTM training.

[0192] Continue to iterate to obtain the updated position. If the new position is better than the old position, update the old position; continuously iterate and calculate, and output the best fitness value and sparrow individuals.

[0193] Step S4.9: Construct an optimized pre-trained LSTM neural network model with the optimal maximum number of training times, the optimal number of hidden units, and the optimal initial learning rate.

[0194] In a specific implementation, set the optimization objectives of the SSA algorithm according to the above Step S4.8 as: the maximum number of training times of the LSTM neural network, the number of hidden units, and the initial learning rate; initialize the number of iterations of the SSA optimization mechanism to 50, the population to 30, the proportion of discoverers to 0.7, the proportion of sparrows aware of danger to 0.2, the warning value to 0.6, and the dimension of the optimized parameters to 3 (the objectives include the maximum number of training times of the LSTM neural network, the number of hidden units, and the initial learning rate), select RMSE as the objective function, and calculate the fitness value according to the objective function; continuously update the positions of the discoverers, followers, and early warning individuals according to Equations (28)-(30) during the calculation; after each optimization is completed, transmit the parameters to the LSTM, calculate the fitness according to the results of the LSTM training, and find the current best fitness individual and the worst fitness individual; continue to iterate to obtain the updated positions, and if the new positions are better than the old positions, update the old positions; continuously iterate and calculate, and output the best fitness value and sparrow individuals. Thus, the SSA optimization mechanism and the LSTM neural network constitute the SSA-LSTM building daily load prediction framework.

[0195] In the case where the determined number of modes is 8, input the dataset reconstructed again after MVMD decomposition into the SSA-LSTM prediction framework for training and verification prediction. It should be noted that the final predicted value is the sum of the predicted values of each mode of the building heat load, and thus the MVMD-SSA-LSTM building daily load prediction result is obtained. Taking December 15th as an example, the comparison chart of the prediction result of this case with that of a common load calculation software is shown in the appendix Figure 11 . Compared with the actual situation, the results using the prediction model can well match the actual daily load value; the prediction result of DEST is on the high side. Although it uses dynamic calculation and is relatively sensitive to changes in external temperature, during periods of drastic temperature changes, it will overestimate the heat loss of the building, and the short-term load fluctuations may be over-amplified; Swell uses static heat balance steady-state calculation, which is suitable for rapid estimation and preliminary design analysis. When considering the building external environment, some dynamic change factors (such as changes in indoor heat sources) may be ignored. In addition, the consideration of the heat storage effect is also insufficient. Therefore, the Swell model may underestimate the actual demand when predicting the heat load.

[0196] Embodiment 2:

[0197] This embodiment proposes a dynamic daily load prediction system for energy-saving heating in public buildings in winter, including: a data acquisition module, a data acquisition module, a data reconstruction module, and a prediction result output module;

[0198] The data acquisition module is connected to the data acquisition module, the data acquisition module is connected to the data reconstruction module, and the data reconstruction module is connected to the prediction result output module;

[0199] The data acquisition module is used to acquire the influencing factors of heat supply to be predicted;

[0200] The modal decomposition module is used to decompose each influencing factor of heat supply to be predicted to obtain multiple modes of each influencing factor;

[0201] The data reconstruction module is used to select the mode with the highest correlation coefficient with the heat load from the multiple modes of each influencing factor, and form the reconstructed data with the influencing factors of heat supply to be predicted;

[0202] The prediction result output module is used to input the reconstructed data into the pre-trained LSTM neural network model to obtain the daily load prediction result. The pre-trained LSTM neural network model is trained with the historical data of the influencing factors of heat supply, and the sparrow search method is used to optimize the parameters of the LSTM neural network model.

[0203] Embodiment 3:

[0204] This embodiment proposes an electronic device, including: one or more processors, and a memory. The memory is used to store instructions. When the instructions are executed by the one or more processors, the one or more processors execute the dynamic daily load prediction method for energy-saving heat supply in winter of public buildings.

[0205] The electronic device can be a mobile phone, a computer, a tablet computer, etc., including a memory and a processor. A computer program is stored on the memory. When the computer program is executed by the processor, it implements the dynamic daily load prediction method for energy-saving heat supply in winter of public buildings as described in the embodiment. It can be understood that the electronic device can also include an input / output (I / O) interface and a communication component.

[0206] Among them, the processor is used to execute all or part of the steps in the dynamic daily load prediction method for energy-saving heat supply in winter of public buildings as described in the above embodiment. The memory is used to store various types of data, which can include, for example, instructions of any application program or method in the electronic device, and data related to the application program.

[0207] The processor may be implemented by an Application Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), a Programmable Logic Device (PLD), a Field Programmable Gate Array (FPGA), a controller, a microcontroller, a microprocessor, or other electronic components, and is used to execute the dynamic daily load prediction method for winter energy-saving heating of public buildings described in the above embodiments.

[0208] Embodiment 4:

[0209] This embodiment provides a computer-readable storage medium storing executable instructions, which, when implemented in the form of a software functional unit and sold or used as an independent product, may be stored in a computer-readable storage medium.

[0210] This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the dynamic daily load prediction method for winter energy-saving heating of public buildings described in various embodiments of the present application.

[0211] The foregoing storage medium includes: flash memory, hard disk, multimedia card, card-type memory (such as SD (Secure Digital Memory Card) or DX (abbreviation for Memory Data Register, MDR), memory data register, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, server, APP (abbreviation for Application, application software), application mall, and other various media that can store program check codes. A computer program is stored thereon, and when the computer program is executed by a processor, it can implement each step of the dynamic daily load prediction method for winter energy-saving heating of public buildings described above.

[0212] Embodiment 5:

[0213] This embodiment provides a computer program product including a computer program or instructions, which, when executed by a processor, implement the dynamic daily load prediction method for winter energy-saving heating of public buildings described above.

[0214] Based on such understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a computer program product.

[0215] Each embodiment in the present application is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments.

[0216] The protection scope of the present application is not limited to the above embodiments. Obviously, those skilled in the art can make various changes and modifications to the present disclosure without departing from the scope and spirit of the present disclosure. If these changes and modifications fall within the scope of the claims of the present disclosure and their equivalent technologies, the intention of the present disclosure also includes these changes and modifications.

Claims

1. A dynamic daily load forecasting method for energy-saving heating in winter for public buildings, characterized in that: include: Obtain the influencing factors of the heat supply to be predicted; Decompose each influencing factor of the heating to be predicted to obtain multiple modes of each influencing factor; Among the multiple modes of each influencing factor, the mode with the highest correlation coefficient with the heat load is selected, and the mode is combined with the influencing factor of the heating to be predicted to form the reconstructed data; The reconstructed data is input into a pre-trained LSTM neural network model to obtain a daily load forecast result. The pre-trained LSTM neural network model is trained using historical data of factors affecting heating, and the parameters of the LSTM neural network model are optimized using a sparrow search method.

2. A dynamic daily load forecasting method for energy-saving heating in winter for public buildings according to claim 1, characterized in that: The training process of the pre-trained LSTM neural network model includes: Obtain historical data on factors affecting heating, and use the heat load of historical buildings as prediction labels; Calculate the Pearson correlation coefficient based on the historical data of the factors affecting heating and the prediction labels, remove the historical data of the factors affecting heating whose Pearson correlation coefficient is less than a preset threshold, and obtain a first historical reconstructed data set; Decomposing the historical data of each influencing factor of heating in the first historical reconstruction data set to obtain multiple modes of the historical data of each influencing factor; For multiple modes of historical data of each influencing factor, calculate the Pearson correlation coefficient and select the corresponding mode with the highest Pearson correlation coefficient; The corresponding mode with the highest Pearson correlation coefficient and the historical data of the influencing factors of heating are combined into a second historical reconstruction data set; The second historical reconstruction data set is used to train the LSTM neural network model to obtain a pre-trained LSTM neural network model.

3. A dynamic daily load forecasting method for energy-saving heating in winter for public buildings according to claim 1, characterized in that: The method of optimizing the parameters of the LSTM neural network model by using the sparrow search method includes: Normalizing the second historical reconstruction data set; Taking the maximum number of training times, the number of hidden units and the initial learning rate in the pre-trained LSTM neural network model as the training target of the sparrow search method, based on the normalized second historical reconstruction data set, the sparrow search method is used to iteratively train the maximum number of training times, the number of hidden units and the initial learning rate in the trained LSTM neural network model to obtain the optimal maximum number of training times, the optimal number of hidden units and the optimal initial learning rate; Construct an optimized pre-trained LSTM neural network model with the optimal maximum number of training times, the optimal number of hidden units, and the optimal initial learning rate.

4. A dynamic daily load forecasting method for energy-saving heating in winter for public buildings according to claim 2, characterized in that: The Pearson correlation coefficient is calculated as follows: Among them, γ X,Y is the Pearson correlation coefficient between the historical data of the factors affecting heating and the predicted labels, m is the number of historical data of the factors affecting heating, X r is the rth historical data of the factors affecting heating, Y r is the rth predicted label, i.e., the heat load of the historical building, is the average value of historical data of factors affecting heating, is the average value of the predicted labels, γ X,Y The value range is [-1,1]. X,Y When it is close to 1 or -1, it indicates that the correlation between the two variables X and Y is higher. X,Y The closer it is to 0, the lower the correlation between the two variables X and Y.

5. A dynamic daily load forecasting method for energy-saving heating in winter for public buildings according to claim 2, characterized in that: The historical data of each heating influencing factor in the first historical reconstruction data set is decomposed to obtain multiple modes of the historical data of each influencing factor, including: Using the historical data of each influencing factor of heating in the first historical reconstruction data set as input data of the data channel, and using multiple modes of the historical data of each influencing factor as output data; According to the relationship between input data and output data, a multivariate modulated oscillation set is obtained; Performing a Hilbert transform on each element in the multivariate modulated oscillation set to obtain a vector analytical signal, wherein the vector analytical signal is composed of a center frequency of the vector analytical signal and a frequency domain representation of the kth mode in the cth channel; Initialize the Lagrange multiplier, the center frequency of the vector analytical signal, and the frequency domain representation of the kth mode in the cth channel; The frequency domain representation of the kth mode in the cth channel is updated using Lagrange multipliers and penalty factors; The center frequency is updated using the frequency domain representation of the kth mode in the cth channel; The Lagrange multiplier is updated using noise tolerance; When the preset convergence condition is reached, the update of the Lagrange multiplier, the center frequency and the frequency domain representation of the kth mode in the cth channel is terminated, and the frequency domain representation of the kth mode in the cth channel after the n+1th iteration is output. The frequency domain representation of the kth mode in the cth channel after the n+1th iteration corresponds to multiple modes of the historical data of each influencing factor after the update.

6. A dynamic daily load forecasting method for energy-saving heating in winter for public buildings according to claim 5, characterized in that: The Lagrange multiplier and the penalty factor are used to update the frequency domain representation of the kth mode in the cth channel, and the calculation formula is as follows: Among them, n is the maximum number of iterations, α is the penalty factor, and λ c (ω) is the Lagrange multiplier, ω is the frequency variable, ω k is the center frequency of the kth mode, x c (ω) is the frequency domain representation of the input data of the cth channel, u k,c (ω) is the frequency domain representation of the kth mode in the cth channel, is the frequency domain representation of the kth mode in the cth channel after the n+1th iteration; The center frequency is updated by using the frequency domain representation of the kth mode in the cth channel, and the calculation formula is as follows: Where c is the number of data channels, is the center frequency of the kth mode at the n+1th iteration; The Lagrange multiplier is updated by adopting the noise tolerance, and the calculation formula is as follows: Where τ is the noise margin, is the Lagrange multiplier of the c-th channel at the n-th iteration.

7. A dynamic daily load forecasting system for energy-saving heating in winter for public buildings, characterized in that: include: A data acquisition module, used to acquire the influencing factors of the heating supply to be predicted; A modal decomposition module is used to decompose each influencing factor of the heating to be predicted to obtain multiple modes of each influencing factor; A data reconstruction module is used to select the mode with the highest correlation coefficient with the heat load from multiple modes of each influencing factor, and to form reconstructed data with the influencing factor of the heating to be predicted; The prediction result output module is used to input the reconstructed data into the pre-trained LSTM neural network model to obtain the daily load prediction result. The pre-trained LSTM neural network model is trained using historical data of factors affecting heating, and the sparrow search method is used to optimize the parameters of the LSTM neural network model.

8. An electronic device, characterized in that: include: One or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the dynamic daily load forecasting method for energy-saving heating in winter for public buildings as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The device stores executable instructions, which, when executed, enable the processor to execute a dynamic daily load forecasting method for energy-saving heating in winter for public buildings as described in any one of claims 1 to 6.

10. A computer program product comprising a computer program or instructions, characterized in that When the computer program or instruction is executed by a processor, a dynamic daily load forecasting method for energy-saving heating in winter for a public building as described in any one of claims 1 to 6 is implemented.

Images