Building air conditioner load prediction method and system

By using time convolution network and improved signal processing algorithms in air conditioner load prediction, the problem of low prediction accuracy under small sample data is solved, and higher prediction accuracy and stability are achieved, suitable for air conditioner load prediction in actual buildings.

CN118964849APending Publication Date: 2024-11-15STATE GRID JIANGSU ELECTRIC POWER CO LTD MARKETING SERVICE CENT +1

Patent Information

Application Number
CN202411046794.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing air conditioner load prediction method has low accuracy under the conditions of small sample data, ignoring the deviation of small sample data, and the traditional method has strict requirements on training data and cannot effectively adapt to the complex changes in actual buildings.

Method used

The time convolution network (TCN) is used to combine the improved variational modal decomposition IVMD algorithm, arrangement entropy PE algorithm and SG filtering algorithm to decompose, denoising and smoothing the historical load signals to build a prediction model suitable for small sample data. Optimize model parameters to improve prediction accuracy.

Benefits of technology

It improves the accuracy and stability of air conditioner load prediction, reduces the impact of noise, is suitable for high-precision prediction of existing buildings, reduces prediction errors, and enhances the generalization ability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118964849A_ABST
    Figure CN118964849A_ABST
Patent Text Reader

Abstract

The invention discloses a building air conditioner load prediction method and system, and the method comprises the steps: collecting the historical load signal and influence factor historical data of a target building air conditioner, and constructing a small sample original data set; data preprocessing: using an improved variational mode decomposition (IVMD) algorithm to deconstruct a complex historical load signal, and combining a permutation entropy PE algorithm and an SG filtering algorithm to reduce noise and extract the signal; a TCN load prediction model is constructed and the model is trained by using sample set data, so that the risks of over-fitting and under-fitting are effectively relieved; and preprocessing the historical load signals of two days before the to-be-predicted day and the historical data of the influence factors, and inputting the preprocessed historical load signals and the preprocessed historical data into the trained TCN load prediction model to obtain an air conditioner load prediction result of the prediction day. According to the method, the problems of low signal decomposition accuracy and small sample data sets can be well solved, the prediction accuracy is improved, and high stability and generalization ability are shown.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for predicting building air conditioning load. Background Art

[0002] As global problems such as energy shortage and environmental pollution become increasingly serious, it is imperative to reduce energy consumption in order to promote the process of high-quality development. The construction industry accounts for about 40% of global energy consumption and greenhouse gas emissions, and more than half of the energy consumption in the construction industry is used for heating, ventilation and air-conditioning systems. In the energy management of the construction industry, air-conditioning load forecasting technology is a key part of building energy management, and it is necessary to further improve the prediction accuracy of air-conditioning load forecasting technology for actual buildings in order to improve energy utilization efficiency.

[0003] At present, the mainstream load forecasting method is the mechanism model forecasting method, but this method is usually used to predict buildings that have not yet been built, and some assumptions must be made about the situation after use. Considering some complex situations that actual buildings may face, such as sudden changes in operating conditions or weather conditions, the resulting scheduling information or external temperature may be significantly different from the predetermined situation in the simulation environment, which will cause a large gap between the prediction and the actual situation after being put into use. In terms of neural network forecasting methods, the existing technology often uses back propagation networks (BP), generative adversarial networks (GAN) and long short-term memory networks (LSTM) to predict air conditioning loads, but these methods have strict requirements on training data and require a large number of training data sets to produce stable prediction results. However, not all occasions have enough data for training. In practice, the variable data selected are mostly short-term load data, and the sample points are very limited. In addition, the data iteration coverage will cause insufficient data volume, resulting in a decrease in the accuracy of the existing model. These methods ignore the impact of small sample data deviation on the prediction results. Therefore, in order to better improve energy efficiency, it is necessary to have an air conditioning load forecasting method that is more in line with reality and more accurate. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a method and system for predicting building air-conditioning load, which can improve the prediction performance of air-conditioning load when the training sample data set is small, and has high reliability and high accuracy.

[0005] The present invention adopts the following technical solution.

[0006] In a first aspect, the present invention provides a method for predicting air conditioning load of a building, the method comprising:

[0007] Collect multiple groups of historical load signals of target building air conditioners and historical data of influencing factors at the same time scale according to preset time intervals to construct a small sample original data set;

[0008] Preprocess the small sample original data set to obtain multiple sets of complete and available data combinations to form a sample set;

[0009] Dividing the sample set into an input sample set and an output sample set according to the sampling time of the data combination;

[0010] Based on the input sample set, the constructed TCN load forecasting model is trained to obtain the forecasting result corresponding to each input sample;

[0011] The preprocessed historical load signal in the output sample set is taken as the true value, and the prediction result at the same prediction time of the corresponding input sample is taken as the predicted value, and the mean square error loss function between the true value and the predicted value is calculated;

[0012] Optimize the parameters of the model with the goal of minimizing the mean square error loss function to obtain a finally trained TCN load forecasting model;

[0013] After preprocessing the historical load signals and historical data of influencing factors two days before the forecast day, they are input into the trained TCN load forecasting model to obtain the air-conditioning load forecast result for the forecast day.

[0014] Optionally, the influencing factors include the dry-bulb temperature, wet-bulb temperature, relative humidity, atmospheric pressure, wind speed, dew point temperature of the target building, lighting power in the target building, inlet and outlet water temperatures of the air-conditioning condensing device, and inlet and outlet water temperatures of the air-conditioning compression device.

[0015] Optionally, the sampling period of the small sample original data set is set to within 1 month, and the sampling time interval is 1 hour.

[0016] Optionally, the preprocessing step includes:

[0017] Calculate the Pearson correlation coefficient between the historical data of each influencing factor and the historical load signal of the corresponding day respectively;

[0018] The historical data of influencing factors with Pearson correlation coefficients higher than the set values ​​on each day are screened out respectively to obtain the corresponding influencing feature data sets for each day.

[0019] Optionally, the step of preprocessing the small sample original data set further includes:

[0020] The improved variational mode decomposition (IVMD) algorithm is used to decompose the daily historical load signals and obtain the corresponding load decomposition sequences for each day.

[0021] Performing noise reduction processing on each of the load decomposition sequences respectively by using a permutation entropy PE algorithm;

[0022] Smoothing each of the de-noised load decomposition sequences using an SG filtering algorithm;

[0023] The load decomposition sequences after the smoothing process are reconstructed respectively to obtain the corresponding load reconstruction signals of each day after the preprocessing.

[0024] Optionally, dividing the sample set into an input sample set and an output sample set according to the sampling time of the data combination includes:

[0025] According to the sampling time, the corresponding relationship between each group of input samples and output samples is set as follows: the impact feature data sets and load signals on the S-1th day and the S-2th day are used as the input samples of the training model, and the load signal corresponding to the Sth day is used as the true value of the output sample in its training model; where S represents the date, S-1 represents the date one day before the Sth day, and S-2 represents the date two days before the Sth day.

[0026] Optionally, the step of using an improved variational mode decomposition (IVMD) algorithm to decompose daily historical load signals comprises:

[0027] Step 1: Initialize the modal center frequency and modal function;

[0028] Step 2: Modal number K after particle swarm optimization s , respectively construct the constrained variational model of each historical load signal:

[0029]

[0030] Where, X s is the historical load sequence collected on the Sth day; μ k is the kth mode function; ω k The center frequency of the kth mode function; δ(t) is the Dirac function, is the Hilbert transform, is the partial differential at time t, is Gaussian smoothing, μ k (t) is the kth decomposition mode at time t, is the modified index, * indicates the convolution operation;

[0031] Step 3: Combine the penalty factor α optimized by the particle swarm algorithm s , the Lagrangian function is used to transform each constrained variational model into an unconstrained model:

[0032]

[0033] In the formula, L({μ k},{ω k},λ) is the Lagrangian function, λ(t) is the Lagrangian multiplier at time t, X s (t) is the historical load sequence X s The signal at time t;

[0034] Step 4: Based on the solution of the unconstrained model, reconstruct the K with the minimum error s The load decomposition sequence corresponding to its historical load signal is obtained by using an intrinsic mode function.

[0035] Optionally, the modal number K s and penalty factor α s The optimization steps include:

[0036] Step A: Initialize the particle population parameters, including: selecting the number of particles; setting the maximum number of iterations; determining the inertia weight and learning factor; defining the search space and setting the modal number K s and penalty factor α s scope;

[0037] Step B: Randomly generate the initial position and initial velocity of each particle in the search space, where the penalty factor α s and the modal number K s The position vectors of each particle: V j =(α s,j ,K s,j ), α s,j represents the penalty factor value of the jth particle, K s,j represents the decomposed modal value of the jth particle;

[0038] Step C: Calculate the fitness of each particle, and determine the individual initial optimal position and population optimal position of each particle according to the fitness;

[0039] Step D: Based on the current position of each particle, the position and velocity of each particle are updated in combination with the inertia weight and the learning factor;

[0040] Step E: Based on the current speed and position of each particle, recalculate the fitness of each particle and update the optimal position of each particle;

[0041] Step F: Repeat steps D to E until the maximum number of iterations is met or the fitness value meets the convergence condition. Select the individual with the highest fitness value in the final population as the optimal solution, and output the modal number K corresponding to the optimal solution. s and penalty factor α s .

[0042] Optionally, the formulas for updating the position and velocity of each particle in combination with the inertia weight and the learning factor are as follows:

[0043] v j (l+1)=β·v j (l)+c1·r1·(pbest j -x j (l))+c2·r2·(gbest j -x j (l))

[0044] x j (l+1)=x j (l)+v j (l+1)

[0045] In the formula, v j (l) and v j (l+1) respectively represent the j particle in its vector V after the lth and l+1th iterations. j =(α s,j ,K s,j ) dimension; x j (l) and x j (l+1) respectively represent the j particle in its vector V after the lth and l+1th iterations. j =(α s,j ,K s,j ) dimension; β represents the inertia weight; c1 and c2 are two different learning factors; r1 and r2 are both random numbers between 0 and 1; l represents the current number of iterations; pbest j represents the individual optimal position of particle j in the lth iteration; pbest j represents the optimal position of the population at the lth iteration.

[0046] Optionally, the step of performing noise reduction processing on each of the load decomposition sequences by using a permutation entropy PE algorithm comprises:

[0047] According to the set embedding dimension m and delay time τ, the load decomposition sequence C of each day is S Each vector in is reconstructed in phase space to obtain the corresponding time reconstruction sequence:

[0048] Based on the time reconstruction sequence of each vector, the permutation entropy of each corresponding vector is calculated;

[0049] Each permutation entropy is normalized to obtain its own H p , H p The vectors with values ​​greater than 0.7 are removed from the load decomposition sequence of the corresponding day to obtain the load decomposition sequence X after noise reduction on the corresponding day. * S .

[0050] Optionally, the step of smoothing each of the noise-reduced load decomposition sequences by using an SG filtering algorithm comprises:

[0051] Step a: Calculate the load decomposition sequence X after noise reduction * S The vector residuals A in i ;

[0052] Step b: Initialize the parameters of the SG filter, including the window size and polynomial order;

[0053] Step c: In the current sliding window, use the least squares method to calculate the current vector residual A i Perform polynomial fitting to obtain the fitted polynomial and smoothing coefficient;

[0054] Step d: Calculate the smoothing value of the window center point based on the fitted polynomial and the smoothing coefficient, and use the smoothing value to replace the original sequence C * S The corresponding value in ;

[0055] Step e: Move the sliding window and repeat steps b to d until all data are replaced to obtain a smoothed signal.

[0056] Optionally, the constructed TCN load forecasting model includes an input layer, multiple hidden layers and an output layer connected in sequence; each of the hidden layers is used to perform causal convolution and dilated convolution on the input data to different degrees;

[0057] The convolution kernel size in the hidden layer of the TCN load forecasting model is 2, and ReLU is selected as the activation function; the network optimizer used in training and optimizing the TCN load forecasting model is Adam.

[0058] Optionally, the expression of the mean square error loss function is as follows:

[0059]

[0060] Where N is the number of samples, y n is the true value of the nth sample, is the predicted value obtained from the nth sample training.

[0061] In a second aspect, the present invention provides a building air conditioning load prediction system, which runs the steps of any method described in the first aspect of the present invention, and the system comprises:

[0062] The data collection module is used to collect multiple groups of historical load signals of target building air conditioners and historical data of influencing factors at the same time scale according to preset time intervals, and construct a small sample original data set;

[0063] The preprocessing module is used to preprocess the small sample original data set to obtain multiple sets of complete and available data combinations to form a sample set;

[0064] A sample division module, used for dividing the sample set into an input sample set and an output sample set according to the sampling time of the data combination;

[0065] A model training module, used to train the constructed TCN load forecasting model based on the input sample set to obtain the forecasting result corresponding to each input sample;

[0066] The error calculation module is used to take the preprocessed historical load signal in the output sample set as the true value, and the prediction result at the same prediction time of the corresponding input sample as the prediction value, and calculate the mean square error loss function between the true value and the prediction value;

[0067] A model optimization module, used to optimize the parameters of the model with the goal of minimizing the mean square error loss function, and obtain a finally trained TCN load forecasting model;

[0068] The load forecasting module is used to pre-process the historical load signals and historical data of influencing factors two days before the forecast day, and input them into the trained TCN load forecasting model to obtain the air conditioning load forecast result for the forecast day.

[0069] In a third aspect, the present invention provides a terminal, including a processor and a storage medium;

[0070] The storage medium is used to store instructions;

[0071] The processor is used to operate according to the instructions to execute the steps of any one of the methods described in the first aspect of the present invention.

[0072] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect of the present invention.

[0073] The beneficial effects of the present invention are as follows:

[0074] 1. The present invention is more suitable for air-conditioning load prediction of existing buildings, avoiding the default assumptions about building operating conditions in traditional mechanism models, reducing the situation where some sudden changes that buildings may face cause reduced prediction accuracy; and can well solve the problems of low signal decomposition accuracy and small sample data sets, improve prediction accuracy, and show high stability and generalization ability.

[0075] 2. The present invention proposes an improved variational mode decomposition (IVMD) algorithm to decompose complex load signals, and uses a particle swarm optimization algorithm to optimize the penalty factor and mode number parameters of the original variational mode decomposition (VMD) method, thereby improving the decomposition accuracy and convergence speed, optimizing the variational mode decomposition algorithm's sensitivity to the decomposed mode number and the difficulty in determining the penalty factor, making it more conducive to noise identification, and further solving the problem of multi-component load signals affecting prediction accuracy.

[0076] 3. The present invention combines the permutation entropy PE algorithm and the SG filtering algorithm to perform noise reduction and smoothing processing on the historical load signal. By identifying and eliminating high-frequency noise components through PE, the noise signal is further eliminated, the recognition ability of nonlinear and non-stationary signals is enhanced, and the influence of small sample deviation data on the prediction results is reduced.

[0077] 4. In view of the problem that data sample points of buildings are very limited in practice, the present invention establishes a TCN load forecasting model based on small sample load data, which reduces the risk of overfitting due to insufficient sample data and ensures stability and prediction accuracy. The method of the present invention is more in line with reality and has higher prediction accuracy, and is suitable for promotion and application. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 This is a flow chart of the building air conditioning load prediction method in the present invention;

[0079] Figure 2 It is a schematic diagram of the structure of the temporal convolutional network TCN in the present invention;

[0080] Figure 3 This is a noise comparison diagram before and after data preprocessing in the present invention;

[0081] Figure 4 A comparison diagram of the actual value and the predicted value of the air conditioning load in the present invention;

[0082] Figure 5 This is a structural principle block diagram of the building air conditioning load prediction system in the present invention. DETAILED DESCRIPTION

[0083] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in the present invention are only part of the embodiments of the present invention, not all of the embodiments. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0084] Embodiment 1:

[0085] Reference Figure 1The embodiment of the present invention provides a method for predicting air conditioning load of a building, which specifically includes the following steps:

[0086] Step S1, collecting multiple groups of historical load signals of target building air conditioners and historical data of influencing factors at the same time scale according to preset time intervals, and constructing a small sample original data set;

[0087] Among them, the factors affecting the air-conditioning load include: the dry-bulb temperature, wet-bulb temperature, relative humidity, atmospheric pressure, wind speed, dew point temperature of the target building, the lighting power in the target building, the inlet and outlet water temperature of the air-conditioning condensing device, the inlet and outlet water temperature of the air-conditioning compression device, and the power of other energy-consuming equipment in the target building.

[0088] In a preferred but non-limiting embodiment, the sampling period of the data in step S1 is set to be within 1 month, the sampling time is from 0:00 to 23:00 every day, and the sampling time interval is 1 hour.

[0089] Step S2, preprocessing the small sample original data set to obtain multiple groups of complete and available data combinations to form a sample set;

[0090] Furthermore, the preprocessing process in step S2 includes:

[0091] S2.1: Calculate the Pearson correlation coefficient between the historical data of each influencing factor and the historical load signal of the corresponding day respectively;

[0092] S2.2: Filter out the historical data of influencing factors with a Pearson correlation coefficient higher than 0.5 on each day, and obtain the corresponding influencing feature data set for each day;

[0093] When the absolute value of the Pearson correlation coefficient is greater than 0.5, it indicates that the correlation of the influencing factor is high. S represents the impact feature dataset on the Sth day after screening;

[0094] S2.3: Use the improved variational mode decomposition (IVMD) algorithm to decompose the daily historical load signals and obtain the corresponding load decomposition sequence for each day;

[0095] This embodiment uses an improved variational mode decomposition (IVMD) algorithm to process the original signal sequence of each day, and uses a particle swarm optimization algorithm (PSO) to optimize the penalty factor α and the mode number K of the original variational mode decomposition (VMD) method to improve the decomposition accuracy and convergence speed; the specific process includes:

[0096] Step 1: Initialize the modal center frequency and modal function;

[0097] Step 2: Modal number K after particle swarm optimization s, respectively construct the constrained variational model of each historical load signal:

[0098]

[0099]

[0100] Where, X s is the historical load sequence collected on the Sth day; μ k is the kth mode function; ω k The center frequency of the kth mode function; δ(t) is the Dirac function, is the Hilbert transform, is the partial differential at time t, is Gaussian smoothing, μ k (t) is the kth decomposition mode at time t, is the modified index, * indicates the convolution operation;

[0101] Step 3: Combine the penalty factor α optimized by the particle swarm algorithm s , the Lagrangian function is used to transform each constrained variational model into an unconstrained model:

[0102]

[0103] In the formula, L({μ k},{ω k},λ) is the Lagrangian function, λ is the Lagrangian multiplier at time t, X s (t) is the original signal sequence X s The signal at time t;

[0104] Step 4: Based on the solution of the unconstrained model, reconstruct the K with the minimum error s The load decomposition sequence corresponding to its historical load signal is obtained by using the intrinsic mode function; that is, X c S =[x 1,s ,x 2,s ,......,x k,s ].

[0105] As an embodiment of the present invention, the modal number K in step 2 s and the penalty factor α in step 3 s The optimization steps include:

[0106] Step A: Initialize the particle population parameters, including: selecting the number of particles; setting the maximum number of iterations; determining the inertia weight and learning factor; defining the search space and setting the modal number K s and penalty factor α s scope;

[0107] Step B: Randomly generate the initial position and initial velocity of each particle in the search space, where the penalty factor α s and the modal number K s The position vectors of each particle: V j =(α s,j ,K s,j ), α s,j represents the penalty factor value of the jth particle, K s,j represents the decomposed modal value of the jth particle;

[0108] Step C: Calculate the fitness of each particle, and determine the individual initial optimal position and population optimal position of each particle according to the fitness;

[0109] Step D: Based on the current position of each particle, the position and velocity of each particle are updated in combination with the inertia weight and the learning factor. The update formulas are as follows:

[0110] v j (l+1)=β·v j (l)+c1·r1·(pbest j -x j (l))+c2·r2·(gbest j -x j (l))

[0111] x j (l+1)=x j (l)+v j (l+1)

[0112] In the formula, v j (l) and v j (l+1) respectively represent the j particle in its vector V after the lth and l+1th iterations. j =(α s,j ,K s,j ) dimension; x j (l) and x j (l+1) respectively represent the j particle in its vector V after the lth and l+1th iterations. j =(α s,j ,K s,j ) dimension; β represents the inertia weight; c1 and c2 are two different learning factors; r1 and r2 are both random numbers between 0 and 1; l represents the current number of iterations; pbest j represents the individual optimal position of particle j in the lth iteration; pbest j represents the optimal position of the population at the lth iteration.

[0113] Step E: Based on the current speed and position of each particle, recalculate the fitness of each particle and update the optimal position of each particle;

[0114] Step F: Repeat steps D to E until the maximum number of iterations is met or the fitness value meets the convergence condition. Select the individual with the highest fitness value in the final population as the optimal solution, and output the modal number K corresponding to the optimal solution. s and penalty factor α s .

[0115] S2.4: performing noise reduction processing on each of the load decomposition sequences by using the permutation entropy PE algorithm; specifically comprising:

[0116] Step a: According to the set embedding dimension m and delay time τ, the load decomposition sequence X of each day is c S =[x 1,s ,x 2,s ,......,x k,s ] is reconstructed in phase space to obtain the corresponding time reconstruction sequence:

[0117] x i,s '=[x i,s (t),x i,s (t+τ),x i,s (t+2τ),......,x i,s (t+(m-1)τ)]

[0118] In the formula, x i,s ' represents the load decomposition sequence C on the Sth day S =[x 1,s ,x 2,s ,......,x k,s ] is the i-th vector x in i,s The time reconstruction sequence of; i = 1, 2…k;

[0119] Step b: Sort the time reconstruction sequence obtained in step a in ascending order to obtain:

[0120] Q(g)=[x1,x2......,x m ]

[0121] Where g = 1, 2…Z, Z ≤ m!; Q(g) has m! different permutations, and the probability of ascending order is one of m!, calculated as the probability of a symbol sequence appearing P g ,but

[0122] Step c: Based on the probability calculated in step b, calculate the load decomposition sequence X by the following formula:c S The permutation entropy of each vector in is:

[0123]

[0124] Step d: Normalize each of the permutation entropies to obtain their respective H p , H p Vectors with values ​​greater than 0.7 are taken from the load decomposition sequence X c S Remove it and get the load decomposition sequence X after noise reduction * S .

[0125] S2.5: Smoothing each of the de-noised load decomposition sequences using the SG filtering algorithm; specifically including:

[0126] Step a: Calculate the load decomposition sequence X after noise reduction * S The vector residuals A in i ;

[0127] Step b: Initialize the parameters of the SG filter, including the window size and polynomial order;

[0128] Step c: In the current sliding window, use the least squares method to calculate the current vector residual A i Perform polynomial fitting to obtain the fitted polynomial and smoothing coefficient;

[0129] Step d: Calculate the smoothing value of the window center point based on the fitted polynomial and the smoothing coefficient, and use the smoothing value to replace the original sequence C * S The corresponding value in ;

[0130] Step e: Move the sliding window and repeat steps b to d until all data are replaced to obtain a smoothed signal.

[0131] S2.6: Reconstruct the load decomposition sequences after the smoothing process respectively to obtain the corresponding load reconstruction signals of each day after preprocessing:

[0132]

[0133] In the formula, Represents the load reconstruction signal on the Sth day.

[0134] Step S3, dividing the sample set into an input sample set and an output sample set according to the sampling time of the data combination;

[0135] A preferred but non-limiting embodiment, the present invention sets the correspondence between each group of input samples and output samples according to the sampling time: the impact feature data sets and load signals of the S-1th day and the S-2th day are used together as the input samples of the training model, and the load signal corresponding to the Sth day is used as the true value of the output sample in its training model; wherein S represents the date, S-1 represents the date one day before the Sth day; S-2 represents the date two days before the Sth day.

[0136] Step S4: training the constructed TCN load forecasting model based on the input sample set to obtain a forecast result corresponding to each input sample;

[0137] Reference Figure 2 The TCN load forecasting model constructed in this embodiment includes an input layer, multiple hidden layers and an output layer connected in sequence; each of the hidden layers is used to perform causal convolution and dilated convolution on the input data to different degrees; the expansion coefficient of the TCN load forecasting model is [1,2,4,8], the convolution kernel size in the hidden layer is 2, and ReLU is selected as the activation function; the network optimizer used in training and optimizing the TCN load forecasting model is Adam, the dropout factor is 0.05, the learning rate is 0.01, and the number of iterations is set to 150 times.

[0138] Step S5: Take the preprocessed historical load signal in the output sample set as the true value, take the prediction result of the corresponding input sample at the same prediction time as the prediction value, and calculate the mean square error loss function of the true value and the prediction value:

[0139]

[0140] Where N is the number of samples, y n is the true value of the nth sample, is the predicted value obtained from the nth sample training.

[0141] Step S6, optimizing the parameters of the model with the goal of minimizing the mean square error loss function to obtain a finally trained TCN load forecasting model;

[0142] Step S7: After preprocessing the historical load signals and historical data of influencing factors two days before the forecast day, they are input into the trained TCN load forecasting model to obtain the air conditioning load forecast result for the forecast day.

[0143] The building air conditioning load prediction method provided by the present invention is further illustrated below through specific examples.

[0144] The energy consumption data used in this experiment are all from an office building in Shanghai, China. Historical data of various factors affecting air-conditioning load and historical load data are collected. The data set includes internal energy consumption data of an office building located in Huangpu District, Shanghai, China, and corresponding meteorological data collected by the Shanghai Meteorological Bureau. The daily sampling time is from 0:00 to 23:00, the sampling interval is 1 hour, and a total of 744 different data points are collected. The air-conditioning system type is a central air-conditioning system. The distances between the office building and the official meteorological station are approximately 10 kilometers and 5 kilometers respectively, ensuring the relevance and credibility of the meteorological data; the collected data are preprocessed by feature screening, raw data decomposition, noise reduction, and reconstruction; the preprocessed data is trained using a temporal convolutional network to establish a TCN air-conditioning load prediction model based on a small sample; this embodiment selects the data of the day to be predicted, the day before the day to be predicted, and the two days before the day to be predicted as the test set, and the remaining days are used as the training set to support the construction of the corresponding prediction model.

[0145] Furthermore, the collected historical load data is processed by IVMD and PE-SG algorithms. Figure 3 As shown, it can be seen that the signal processed by the method of the present invention can effectively reduce noise interference, which is beneficial to reduce the adverse effects of hardware equipment abnormalities, thereby ensuring prediction accuracy. The processed signal is relatively smooth, and it is easier to identify peaks, thereby improving data quality and feature performance.

[0146] In the experiment, the historical load data of day S and day S-1 processed by the improved variational mode decomposition IVMD algorithm and PE-SG algorithm were used. and Highly correlated feature dataset M with historical load data on day S and day S-1 s 、M s-1 As input to the trained model, the predicted value for day S+1 is obtained through the model output The specific prediction results are as follows: Figure 4 As shown, the upper part represents the comparison between the true value and the predicted value, and the lower part represents the hourly error between the predicted value and the true value; it can be seen from the upper part that the method proposed in the present invention provides better accuracy in predicting the change of cooling load morphology; it can be seen from the lower part that the maximum prediction error of the method of the present invention is within 0.4MW, which effectively reduces the prediction error.

[0147] The following is a comparative experiment of the prediction performance of the method provided by the present invention and the commonly used back propagation (BP) neural network, long short-term memory network (LSTM) and other methods in the prior art by using a variety of evaluation indicators. The indicators used are: mean absolute error (MAE), root mean square error (RMSE), coefficient of variation of root mean square error (CV(RMSE)), determination coefficient (R 2); The comparison results are shown in Table 1, where the TCN in the table is the result of load history data being predicted using only the TCN time convolutional network without undergoing the preprocessing steps such as the variational mode decomposition IVMD algorithm and the PE-SG algorithm provided by the present invention.

[0148] Table 1

[0149] method MAE / (KW) RMSE / (KW) CV(RMSE) / % <![CDATA[R 2 ]]> Method of the present invention 116.38 142.36 25.50 0.95 BP 176.95 247.67 44.36 0.84 LSTM 122.77 191.48 34.29 0.91 TCN 123.93 180.79 32.38 0.92

[0150] The results in Table 1 show that the mean absolute error (MAE), root mean square error (RMSE) and coefficient of variation of the root mean square error (CV(RMSE)) of the method proposed in the present invention are lower than those of other commonly used methods, and the determination coefficient (R2) is higher than that of other commonly used methods. This shows that the building air-conditioning load prediction method proposed in the present invention has good prediction efficiency, good reliability and high accuracy. Compared with the commonly used methods, it has been improved in indicators and better meets the needs of practical applications.

[0151] The beneficial effects of the present invention are as follows:

[0152] 1. The present invention is more suitable for air-conditioning load prediction of existing buildings, avoiding the default assumptions about building operating conditions in traditional mechanism models, reducing the situation where some sudden changes that buildings may face cause reduced prediction accuracy; and can well solve the problems of low signal decomposition accuracy and small sample data sets, improve prediction accuracy, and show high stability and generalization ability.

[0153] 2. The present invention proposes an improved variational mode decomposition (IVMD) algorithm to decompose complex load signals, and uses a particle swarm optimization algorithm to optimize the penalty factor and mode number parameters of the original variational mode decomposition (VMD) method, thereby improving the decomposition accuracy and convergence speed, optimizing the variational mode decomposition algorithm's sensitivity to the decomposed mode number and the difficulty in determining the penalty factor, making it more conducive to noise identification, and further solving the problem of multi-component load signals affecting prediction accuracy.

[0154] 3. The present invention combines the permutation entropy PE algorithm and the SG filtering algorithm to perform noise reduction and smoothing processing on the historical load signal. By identifying and eliminating high-frequency noise components through PE, the noise signal is further eliminated, the recognition ability of nonlinear and non-stationary signals is enhanced, and the influence of small sample deviation data on the prediction results is reduced.

[0155] 4. In view of the problem that data sample points of buildings are very limited in practice, the present invention establishes a TCN load forecasting model based on small sample load data, which reduces the risk of overfitting due to insufficient sample data and ensures stability and prediction accuracy. The method of the present invention is more in line with reality and has higher prediction accuracy, and is suitable for promotion and application.

[0156] Embodiment 2:

[0157] like Figure 5 As shown, the present invention provides a building air conditioning load prediction system, the system is used to implement the steps of the method in the above embodiment 1, and the system specifically includes:

[0158] The data collection module is used to collect multiple groups of historical load signals of target building air conditioners and historical data of influencing factors at the same time scale according to preset time intervals, and construct a small sample original data set;

[0159] The preprocessing module is used to preprocess the small sample original data set to obtain multiple sets of complete and available data combinations to form a sample set;

[0160] A sample division module, used for dividing the sample set into an input sample set and an output sample set according to the sampling time of the data combination;

[0161] A model training module, used to train the constructed TCN load forecasting model based on the input sample set to obtain the forecasting result corresponding to each input sample;

[0162] The error calculation module is used to take the preprocessed historical load signal in the output sample set as the true value, and the prediction result at the same prediction time of the corresponding input sample as the prediction value, and calculate the mean square error loss function between the true value and the prediction value;

[0163] A model optimization module, used to optimize the parameters of the model with the goal of minimizing the mean square error loss function, and obtain a finally trained TCN load forecasting model;

[0164] The load forecasting module is used to pre-process the historical load signals and historical data of influencing factors two days before the forecast day, and input them into the trained TCN load forecasting model to obtain the air conditioning load forecast result for the forecast day.

[0165] The building air conditioning load prediction system provided in the embodiment of the present invention and the building air conditioning load prediction method provided in Example 1 are based on the same technical concept, and can produce the beneficial effects described in Example 1. For the contents not described in detail in this embodiment, please refer to Example 1.

[0166] Embodiment three:

[0167] A terminal provided by an embodiment of the present invention includes a processor and a storage medium;

[0168] The storage medium is used to store instructions;

[0169] The processor is used to operate according to the instruction to execute the steps of the method according to any one of the first embodiments.

[0170] Embodiment 4:

[0171] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of any method described in Embodiment 1 are implemented.

[0172] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0173] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0174] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0175] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting building air conditioning load, characterized in that: Methods include: Collect multiple groups of historical load signals of target building air conditioners and historical data of influencing factors at the same time scale according to preset time intervals to construct a small sample original data set; Preprocess the small sample original data set to obtain multiple sets of complete and available data combinations to form a sample set; Dividing the sample set into an input sample set and an output sample set according to the sampling time of the data combination; Based on the input sample set, the constructed TCN load forecasting model is trained to obtain the forecasting result corresponding to each input sample; The preprocessed historical load signal in the output sample set is taken as the true value, and the prediction result at the same prediction time of the corresponding input sample is taken as the predicted value, and the mean square error loss function between the true value and the predicted value is calculated; Optimize the parameters of the model with the goal of minimizing the mean square error loss function to obtain a finally trained TCN load forecasting model; After preprocessing the historical load signals and historical data of influencing factors two days before the forecast day, they are input into the trained TCN load forecasting model to obtain the air-conditioning load forecast result for the forecast day.

2. The building air conditioning load prediction method according to claim 1, characterized in that: The influencing factors include the dry-bulb temperature, wet-bulb temperature, relative humidity, atmospheric pressure, wind speed, dew point temperature of the target building, lighting power in the target building, inlet and outlet water temperature of the air-conditioning condensing device, and inlet and outlet water temperature of the air-conditioning compression device.

3. The building air conditioning load prediction method according to claim 1, characterized in that: The sampling period of the small sample original data set is set to be within 1 month, and the sampling time interval is 1 hour.

4. The building air conditioning load prediction method according to claim 1, characterized in that: The pre-processing steps include: Calculate the Pearson correlation coefficient between the historical data of each influencing factor and the historical load signal of the corresponding day respectively; The historical data of influencing factors with Pearson correlation coefficients higher than the set values ​​on each day are screened out respectively to obtain the corresponding influencing feature data sets for each day.

5. The building air conditioning load prediction method according to claim 4, characterized in that: The step of preprocessing the small sample original data set also includes: The improved variational mode decomposition (IVMD) algorithm is used to decompose the daily historical load signals and obtain the corresponding load decomposition sequences for each day. Performing noise reduction processing on each of the load decomposition sequences respectively by using a permutation entropy PE algorithm; Smoothing each of the de-noised load decomposition sequences using an SG filtering algorithm; The load decomposition sequences after the smoothing process are reconstructed respectively to obtain the corresponding load reconstruction signals of each day after the preprocessing.

6. The building air conditioning load prediction method according to claim 5, characterized in that: The step of dividing the sample set into an input sample set and an output sample set according to the sampling time of the data combination comprises: According to the sampling time, the corresponding relationship between each group of input samples and output samples is set as follows: the impact feature data sets and load signals on the S-1th day and the S-2th day are used as the input samples of the training model, and the load signal corresponding to the Sth day is used as the true value of the output sample in its training model; where S represents the date, S-1 represents the date one day before the Sth day, and S-2 represents the date two days before the Sth day.

7. The building air conditioning load prediction method according to claim 5, characterized in that: The steps of using the improved variational mode decomposition (IVMD) algorithm to decompose the daily historical load signals include: Step 1: Initialize the modal center frequency and modal function; Step 2: Modal number K after particle swarm optimization s , respectively construct the constrained variational model of each historical load signal: Where, X s is the historical load sequence collected on the Sth day; μ k is the kth mode function; ω k The center frequency of the kth mode function; δ(t) is the Dirac function, is the Hilbert transform, is the partial differential at time t, is Gaussian smoothing, μ k (t) is the kth decomposition mode at time t, is the modified index, * indicates the convolution operation; Step 3: Combine the penalty factor α optimized by the particle swarm algorithm s , the Lagrangian function is used to transform each constrained variational model into an unconstrained model: In the formula, L({μ k },{ω k },λ) is the Lagrangian function, λ(t) is the Lagrangian multiplier at time t, X s (t) is the historical load sequence X s The signal at time t; Step 4: Based on the solution of the unconstrained model, reconstruct the K with the minimum error s The load decomposition sequence corresponding to its historical load signal is obtained by using an intrinsic mode function.

8. The building air conditioning load prediction method according to claim 7, characterized in that: The modal number K s and penalty factor α s The optimization steps include: Step A: Initialize the particle population parameters, including: selecting the number of particles; setting the maximum number of iterations; determining the inertia weight and learning factor; defining the search space and setting the modal number K s and penalty factor α s scope; Step B: Randomly generate the initial position and initial velocity of each particle in the search space, where the penalty factor α s and the modal number K s The position vectors of each particle: V j =(α s,j ,K s,j ), α s,j represents the penalty factor value of the jth particle, K s,j represents the decomposed modal value of the jth particle; Step C: Calculate the fitness of each particle, and determine the individual initial optimal position and population optimal position of each particle according to the fitness; Step D: Based on the current position of each particle, the position and velocity of each particle are updated in combination with the inertia weight and the learning factor; Step E: Based on the current speed and position of each particle, recalculate the fitness of each particle and update the optimal position of each particle; Step F: Repeat steps D to E until the maximum number of iterations is met or the fitness value meets the convergence condition. Select the individual with the highest fitness value in the final population as the optimal solution, and output the modal number K corresponding to the optimal solution. s and penalty factor α s .

9. The building air conditioning load prediction method according to claim 7, characterized in that: The formulas for updating the position and velocity of each particle by combining the inertia weight and the learning factor are as follows: v j (l+1)=β·v j (l)+c1·r1·(pbest j -x j (l))+c2·r2·(gbest j -x j (L)) x j (l+1)=x j (l)+v j (l+1) In the formula, v j (l) and v j (l+1) respectively represent the j particle in its vector V after the lth and l+1th iterations. j =(α s,j ,K s,j ) dimension; x j (l) and x j (l+1) respectively represent the j particle in its vector V after the lth and l+1th iterations. j =(α s,j ,K s,j ) dimension; β represents the inertia weight; c1 and c2 are two different learning factors; r1 and r2 are both random numbers between 0 and 1; l represents the current number of iterations; pbest j represents the individual optimal position of particle j in the lth iteration; pbest j represents the optimal position of the population at the lth iteration.

10. The building air conditioning load prediction method according to claim 5, characterized in that: The step of performing noise reduction processing on each of the load decomposition sequences by using the permutation entropy PE algorithm comprises: According to the set embedding dimension m and delay time τ, the load decomposition sequence C of each day is S Each vector in is reconstructed in phase space to obtain the corresponding time reconstruction sequence: Based on the time reconstruction sequence of each vector, the permutation entropy of each corresponding vector is calculated; Each permutation entropy is normalized to obtain its own H p , H p The vectors with values ​​greater than 0.7 are removed from the load decomposition sequence of the corresponding day to obtain the load decomposition sequence X after noise reduction on the corresponding day. * S .

11. The building air conditioning load prediction method according to claim 10, characterized in that: The step of smoothing each of the noise-reduced load decomposition sequences by using the SG filtering algorithm comprises: Step a: Calculate the load decomposition sequence X after noise reduction * S The vector residuals A in i ; Step b: Initialize the parameters of the SG filter, including the window size and polynomial order; Step c: In the current sliding window, use the least squares method to calculate the current vector residual A i Perform polynomial fitting to obtain the fitted polynomial and smoothing coefficient; Step d: Calculate the smoothing value of the window center point based on the fitted polynomial and smoothing coefficient, and use the smoothing value to replace the original sequence C * S The corresponding value in ; Step e: Move the sliding window and repeat steps b to d until all data are replaced to obtain a smoothed signal.

12. The building air conditioning load prediction method according to claim 10, characterized in that: The constructed TCN load forecasting model includes an input layer, a plurality of hidden layers and an output layer connected in sequence; each of the hidden layers is used to perform causal convolution and dilated convolution on the input data to different degrees; The convolution kernel size in the hidden layer of the TCN load forecasting model is 2, and ReLU is selected as the activation function; the network optimizer used in training and optimizing the TCN load forecasting model is Adam.

13. The building air conditioning load prediction method according to claim 1, characterized in that: The expression of the mean square error loss function is as follows: Where N is the number of samples, y n is the true value of the nth sample, is the predicted value obtained from the nth sample training.

14. A building air conditioning load prediction system, running the building air conditioning load prediction method according to any one of claims 1 to 8, characterized in that: The system includes: The data collection module is used to collect multiple groups of historical load signals of target building air conditioners and historical data of influencing factors at the same time scale according to preset time intervals, and construct a small sample original data set; The preprocessing module is used to preprocess the small sample original data set to obtain multiple sets of complete and available data combinations to form a sample set; A sample division module, used for dividing the sample set into an input sample set and an output sample set according to the sampling time of the data combination; A model training module, used to train the constructed TCN load forecasting model based on the input sample set to obtain the forecasting result corresponding to each input sample; The error calculation module is used to take the preprocessed historical load signal in the output sample set as the true value, and the prediction result at the same prediction time of the corresponding input sample as the prediction value, and calculate the mean square error loss function between the true value and the prediction value; A model optimization module, used to optimize the parameters of the model with the goal of minimizing the mean square error loss function, and obtain a finally trained TCN load forecasting model; The load forecasting module is used to pre-process the historical load signals and historical data of influencing factors two days before the forecast day, and input them into the trained TCN load forecasting model to obtain the air conditioning load forecast result for the forecast day.

15. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-14.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 14 are implemented.

Citation Information

Patent Citations

  • Short-term load prediction method of variational mode decomposition and deep belief network

    CN107392364A

  • Power equipment anomaly detection method and system based on variational mode decomposition

    CN112036042A

  • Short-term building power load prediction method and system based on intelligent electric meter data

    CN112734137A

  • Generalized demand estimation for automated forecasting systems

    US20230186331A1

Cited By

  • Ecological environment influence assessment method and system for mechanical construction of power transmission line

    CN119831131A

  • Power load prediction method and device, electronic equipment and storage medium

    CN119834201A