A method for air conditioning system load forecasting based on deep learning

Abstract

The present invention discloses a method for predicting the load of an air-conditioning system based on deep learning, including step 1, selecting characteristic elements that affect the load of the air-conditioning system; step 2, collecting the original data of the load of the air-conditioning system; step 3, processing the load data; step 4, building a deep learning model based on a neural network; step 5, determining the optimal model of deep learning; step 6, predicting the load of the air-conditioning system. The present invention makes full use of the multi-source characteristic information that affects the energy consumption of the air-conditioning system, and determines the optimal model of deep learning by selecting different numbers of neurons and hidden layers and comparing the prediction results. At the same time, by introducing the Gaussian function to optimize the learning rate #imgabs0# of back propagation, the problem of gradient disappearance that is prone to occur in the calculation process can be solved, the load prediction accuracy can be improved, and energy consumption fluctuation information can be provided, and the operation strategies of various equipment in the refrigeration system can be adjusted in advance, avoiding the energy consumption loss of the air-conditioning system caused by the hysteresis of feedback signal transmission.

Description

Technical Field

[0001] The present invention relates to the field of air conditioning system load prediction, and in particular to an air conditioning system load prediction method based on deep learning. Background Art

[0002] In the past few decades, due to the rapid economic growth, the world's energy consumption and related carbon dioxide emissions have increased dramatically. As an important consumer, the construction industry accounts for about 40% of the world's total energy consumption, of which HVAC systems account for 40% to 50% of building energy consumption, with great energy-saving potential. As an important part of the energy-saving work of HVAC systems, energy consumption forecasting can provide system personnel with energy consumption fluctuation information, help them adjust the operation strategy of chillers, water pumps, and cooling tower equipment in the refrigeration system in advance, and avoid energy consumption losses of the air-conditioning system due to the hysteresis of feedback signal transmission.

[0003] In the era when electronic equipment was not well developed, load forecasting mainly relied on people to manually draw load forecasting curves based on raw data, which added many subjective factors and made it difficult to achieve a high level of accuracy. In recent years, with the development of the computer industry, especially the rapid progress of artificial intelligence technology, the application of machine learning algorithms such as support vector machines and neural network technology to air conditioning load forecasting has become a topic of concern for many researchers, and has also achieved certain results.

[0004] However, in actual projects, when using traditional machine learning algorithms to predict air conditioning loads, they often only use shallow information in the data and fail to fully tap into the multi-source information features that affect the load of the air conditioning system, such as solar radiation intensity, wind speed, season, weekends, holidays, etc. At the same time, the complexity, computational complexity, and accuracy of prediction results of traditional models need to be improved. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an air conditioning system load prediction method based on deep learning in view of the deficiencies of the above-mentioned prior art. The air conditioning system load prediction method based on deep learning fully exploits the multi-source feature information that affects the energy consumption of the air conditioning system. In addition, the present invention also combines the distribution characteristics of the Gaussian function in the deep neural network model, and further optimizes the weight update of the back propagation process, which helps to solve the problem of complex calculation of the deep neural network model and the gradient disappearance problem that is easy to occur in the calculation process. This method simplifies the load prediction model and improves the load prediction accuracy.

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

[0007] A method for air conditioning system load prediction based on deep learning includes the following steps.

[0008] Step 1: Select characteristic elements that affect the load of the air-conditioning system: By analyzing the air-conditioning system to be predicted, m characteristic elements that affect the load of the air-conditioning system are selected.

[0009] Step 2, collecting the original load data of the air-conditioning system: within the set time period and at the set time intervals, the original load data of the air-conditioning system to be predicted is collected, and each collected original load data of the air-conditioning system must include m characteristic elements.

[0010] Step 3, load data processing: each air conditioning system load original data collected in step 2 is dimensionless, so that the m characteristic elements included in each air conditioning system load original data are all formed into dimensionless data; then, 80% of the air conditioning system load original data collected after dimension removal is used as a training set; and 20% of the air conditioning system load original data collected after dimension removal is used as a test set.

[0011] Step 4: Build a deep learning model based on neural network: The deep learning model includes an input layer, y hidden layers and an output layer; the input layer has p neurons, corresponding to p feature elements respectively; p and y are unknown variables to be determined, and p≤m, y≤m; the sigmoid activation function is used as the connection function between layers; the output layer has one neuron, corresponding to the predicted value of the air-conditioning system load at the next moment.

[0012] Step 5, determine the optimal deep learning model: by selecting different p and y values, different deep learning models are obtained; the training set data in step 3 is used to train each deep learning model, and then the connection weight value in each deep learning model is determined; then, the test set data in step 3 is used to verify the model, and different air-conditioning system load prediction values ​​are obtained; by comparing or evaluating the obtained different air-conditioning system load prediction values, the optimal deep learning model is determined; the number of input layer neurons in the optimal deep learning model is p=p1.

[0013] Step 6, air conditioning system load prediction: obtain p1 characteristic elements of the air conditioning system to be predicted at the current moment and perform dimensionless processing, substitute the dimensionless p1 characteristic elements into the deep learning optimal model determined in step 5, and perform air conditioning system load prediction. The output layer result is the predicted value of the air conditioning system load at the next moment.

[0014] In step 1, m=9, and the 9 characteristic elements are historical load value, temperature, humidity, air pressure, wind speed, season, solar radiation intensity, weekends and holidays.

[0015] In step 3, the method of de-dimensionalization is: first, each characteristic element is divided into grades, and each grade corresponds to a dimensionless number; then, each acquired characteristic element is matched with the corresponding grade, and the corresponding dimensionless data is obtained.

[0016] In step 4, the expression of the sigmoid activation function is:

[0017]

[0018] In the formula, x is the predicted output value of the neuron, and S(x) is the output value of the neuron after nonlinear transformation.

[0019] In step 5, the specific method for determining the optimal deep learning model includes the following steps.

[0020] Step 5A, determine the key characteristic elements: first sort the m characteristic elements selected in step 1 from large to small according to their impact on the load of the air conditioning system; then, take the first p0 characteristic elements as the key characteristic elements; wherein the calculation formula of p0 is:

[0021] Step 5B, establish a deep learning model: select different p and y values ​​to obtain different deep learning models; where p∈[p0,m], y∈[y0,m], and y0=2; at this time, the p neurons in the input layer of the deep learning model include at least the p0 key feature elements determined in step 5A.

[0022] Step 5C, calculate the connection weight value: use the training set data in step 3 to train each deep learning model established in step 5B, and then determine the connection weight value in each deep learning model.

[0023] Step 5D, verify the deep learning model: Use the test set data in step 3 to verify each deep learning model with a certain connection weight value to obtain different air conditioning system load prediction values.

[0024] Step 5E, determine the optimal deep learning model: compare or evaluate the load prediction values ​​of different air-conditioning systems obtained in step 5D, and then obtain the optimal deep learning model.

[0025] In step 5E, the MAE indicator is used to evaluate the load forecast values ​​of different air-conditioning systems obtained in step 5D. When the MAE is less than 1%, the deep learning model corresponding to it is the optimal deep learning model.

[0026] The calculation formula of MAE is:

[0027]

[0028] Where, N is the number of original data of air conditioning system load selected from the test set data in step 5D; i is the i-th original data of air conditioning system load selected from the test set data; Y i is the original measured value of the air conditioning system load at the next moment in the test set data; P i It is the air conditioning system load prediction value obtained by using the deep learning model to predict the i-th air conditioning system load original data in step 5D.

[0029] In step 5C, while each deep learning model established in step 5B is trained using the training set data in step 3, the connection weight value of each layer is updated based on the gradient descent method.

[0030] In step 5C, when updating the connection weight value of each layer based on the gradient descent method, the learning rate α(t) in the gradient descent method is designed according to the distribution characteristics of the Gaussian function, and the specific calculation formula is:

[0031]

[0032] Where t represents the number of learning iterations, and h and q are constants.

[0033] The present invention has the following beneficial effects:

[0034] 1. The present invention can comprehensively obtain the influencing factors of the air-conditioning system load, including historical load values, temperature, humidity, air pressure, wind speed, season, solar radiation intensity, weekends and holidays, etc., and unify the dimensions of the variables.

[0035] 2. The deep learning optimal model established by the present invention has multiple hidden layers to form a deep neural network to increase the expressive power of the model. The number of neurons in the hidden layer and the number of hidden layers are finally determined by selecting different numbers of neurons in training and comparing the prediction results.

[0036] 3. The present invention optimizes the learning rate α by introducing a Gaussian function. For deep neural networks with high complexity that require multiple iterations, the algorithm execution efficiency is greatly improved. At the same time, this method helps to solve the gradient vanishing problem that is prone to occur in the calculation process of deep neural network models, simplifies the load forecasting model, and improves the load forecasting accuracy.

[0037] 4. The present invention provides energy consumption fluctuation information to system operation and maintenance personnel, helping them to adjust the operation strategy of each device in the refrigeration system in advance, and avoiding energy consumption loss of the air-conditioning system caused by the hysteresis of feedback signal transmission. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of a method for air conditioning system load prediction based on deep learning in the present invention.

[0039] Figure 2 Schematic diagram of the neural network-based deep learning model constructed for the present invention.

[0040] Figure 3 This is a flowchart of the calculation principle of a single-layer perceptron in a neural network. DETAILED DESCRIPTION

[0041] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific preferred embodiments.

[0042] like Figure 1 As shown, a method for air-conditioning system load prediction based on deep learning includes the following steps.

[0043] Step 1: Select the characteristic elements that affect the air conditioning system load

[0044] By analyzing the air conditioning system to be predicted, m characteristic elements that affect the load of the air conditioning system are selected. In this embodiment, m=9 is preferred. The 9 characteristic elements are historical load value, temperature, humidity, air pressure, wind speed, season, solar radiation intensity, weekends and holidays, etc. Of course, due to the different environments of the air conditioning system to be predicted, the characteristic elements can be adjusted and changed as needed, which is also within the protection scope of this application.

[0045] Step 2: Collect raw data on air conditioning system load

[0046] Within a set time period (such as not less than three years, etc.), the original load data of the air-conditioning system to be predicted is collected at set time intervals (such as 1 hour or 1 day, etc.), and each collected original load data of the air-conditioning system must include m characteristic elements.

[0047] Step 3: Load data processing

[0048] Each of the original data of the air conditioning system load collected in step 2 is de-dimensionalized so that the m characteristic elements included in the original data of the air conditioning system load are all transformed into dimensionless data.

[0049] The preferred method for dimensionless processing is: first, each characteristic element is divided into grades, and each grade corresponds to a dimensionless number; then, each acquired characteristic element is matched with the corresponding grade, and the corresponding dimensionless data is obtained.

[0050] Specifically, when the characteristic element is season, the binning method is:

[0051] During Spring Festival or Autumn, the dimensionless number is 0; during Summer or Winter, the dimensionless number is 1.

[0052] Similarly, if it is a weekend or holiday, the dimensionless number is 1; otherwise, the dimensionless number is 0.

[0053] Of course, according to the environment of the air-conditioning system to be predicted, each characteristic element can also be divided into more levels, which is also within the protection scope of the present application.

[0054] Then, 80% of the original data of air-conditioning system load collected after dimension removal is used as a training set; and 20% of the original data of air-conditioning system load collected after dimension removal is used as a test set.

[0055] Step 4: Build a deep learning model based on neural network

[0056] The deep learning model includes an input layer, y hidden layers and an output layer; the input layer has p neurons, corresponding to p feature elements respectively; p and y are unknown variables to be determined, and p≤m, y≤m; the sigmoid activation function is used as the connection function between layers; the output layer has one neuron, corresponding to the predicted value of the air-conditioning system load at the next moment.

[0057] The expression of the above sigmoid activation function is preferably:

[0058]

[0059] Where x is the predicted output value of the neuron, and S(x) is the output value of the neuron after nonlinear transformation (hereinafter referred to as output value).

[0060] like Figure 2 and 3 As shown in the figure, the activation value and output layer vector of the output layer of the deep learning model are obtained through forward propagation, and the connection weight value and bias value of the deep learning model are recalculated using the back propagation algorithm to iteratively update the deep learning model. The specific calculation process is as follows:

[0061] I j =∑w ij O i +θ j

[0062] In the formula, I j represents the expected output value of node (i.e. neuron) j, w ij represents the connection weight value from node i to node j, O i represents the output value of node i in the previous layer of the computation layer (input value for the input layer), θ j Represents the bias of node j.

[0063]

[0064] O jRepresents the output value of node j and serves as the input value of the next layer.

[0065] Use the back propagation algorithm to recalculate the connection weights and bias values ​​of the prediction model. First:

[0066] For the output layer:

[0067] E rrk =O k (1-O k )(T k -O k )

[0068] In the above formula, E rrk represents the error size generated by computing node k, T k represents the predicted value, O k Represents the output value of node k.

[0069] For the hidden layer:

[0070] E rrj =O j (1-O j )∑ k E rrk w jk

[0071] In the above formula, E rrj represents the error size generated by computing node j, w jk Represents the connection weight value from node j to node k.

[0072] Based on the principle of gradient descent, the connection weight value of each layer is updated:

[0073] Δw ij =(α)E rrj O i

[0074] In the above formula, α represents the learning rate, Δw ij Represents the connection weight error value from node i to node j.

[0075] Therefore, the connection weight value from node i to node j is updated according to the following formula:

[0076] w ij =w ij +Δw ij

[0077] Bias update:

[0078] Δθ j =(α)E rrj

[0079] Therefore, θ jUpdate according to the following formula:

[0080] θ j =θ j +Δθ j

[0081] In the above formula, Δθ j Represents the deviation difference of node j.

[0082] When the value of α is too small, such as α is 0.01, after n iterations, α n When the number of iterations is large, that is, when n takes a large value, α n The value will approach zero. At this time, the back propagation process is prone to gradient vanishing, and the connection weight value and bias update will fall into the local optimal solution.

[0083] Step 5: Determine the optimal model for deep learning. The specific method for determining the optimal model preferably includes the following steps.

[0084] Step 5A: Identify key feature elements

[0085] First, the m characteristic elements selected in step 1 are sorted from large to small according to their impact on the load of the air conditioning system. In this embodiment, the 9 characteristic elements are sorted from large to small as follows: historical load value, temperature, humidity, air pressure, wind speed, season, solar radiation intensity, weekends and holidays.

[0086] Next, the first p0 feature elements are used as key feature elements; the calculation formula for p0 is:

[0087] Step 5B, establish a deep learning model: select different p and y values ​​to obtain different deep learning models; where p∈[p0,m], y∈[y0,m], and y0=2; at this time, the p neurons in the input layer of the deep learning model include at least the p0 key feature elements determined in step 5A.

[0088] The deep learning model corresponding to p0 and y0 is also called the deep learning initial model.

[0089] Step 5C, calculate the connection weight value: use the training set data in step 3 to train each deep learning model established in step 5B, and then determine the connection weight value in each deep learning model.

[0090] In the above training process, the connection weight value of each layer is preferably updated based on the gradient descent method. When updating, the learning rate α in the gradient descent method is preferably designed according to the distribution characteristics of the Gaussian function. The specific calculation formula is:

[0091]

[0092] Where t represents the number of learning iterations, and h and q are constants.

[0093] The present invention optimizes the learning rate α by introducing a Gaussian function. For deep neural networks with high complexity that require multiple loop iterations, the algorithm execution efficiency is greatly improved. At the same time, this method helps to solve the gradient vanishing problem that is prone to occur in the calculation process of deep neural network models.

[0094] Step 5D, verify the deep learning model: Use the test set data in step 3 to verify each deep learning model with a certain connection weight value to obtain different air-conditioning system load prediction values, also known as level 1 load prediction values.

[0095] Step 5E, determine the optimal deep learning model: compare or evaluate the load prediction values ​​of different air-conditioning systems obtained in step 5D, and then obtain the optimal deep learning model.

[0096] In the present invention, the MAE indicator is used to evaluate the load prediction values ​​of different air-conditioning systems obtained in step 5D. When the MAE is less than 1%, the corresponding deep learning model is the optimal deep learning model.

[0097] The calculation formula of the above MAE is preferably:

[0098]

[0099] Where, N is the number of original data of air conditioning system load selected from the test set data in step 5D; i is the i-th original data of air conditioning system load selected from the test set data; Y i is the original measured value of the air conditioning system load at the next moment in the test set data; P i The air conditioning system load forecast obtained by using the deep learning model to predict the i-th air conditioning system load original data in step 5D.

[0100] The number of input layer neurons in the above-mentioned deep learning optimal model is preferably p=p1.

[0101] Step 6, air conditioning system load prediction: obtain p1 characteristic elements of the air conditioning system to be predicted at the current moment and perform dimensionless processing, substitute the dimensionless p1 characteristic elements into the deep learning optimal model determined in step 5, and perform air conditioning system load prediction. The output layer result is the air conditioning system load prediction value or the level 2 load prediction value at the next moment.

[0102] The air conditioning system load prediction value obtained by the present invention can be used to adjust the operation strategy of the chiller, water pump, and cooling tower equipment in the refrigeration system in advance, avoiding energy consumption loss of the air conditioning system caused by the hysteresis of feedback signal transmission.

[0103] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for air conditioning system load prediction based on deep learning, characterized in that: It includes the following steps: Step 1, select characteristic elements that affect the load of the air-conditioning system: by analyzing the air-conditioning system to be predicted, select m=9 characteristic elements that affect the load of the air-conditioning system; The 9 characteristic elements are historical load values, temperature, humidity, air pressure, wind speed, season, solar radiation intensity, weekends and holidays; Step 2, collecting the original load data of the air-conditioning system: within the set time period and at the set time interval, the original load data of the air-conditioning system to be predicted is collected, and each collected original load data of the air-conditioning system must include m characteristic elements; Step 3, load data processing: De-dimensionalize the original data of each air-conditioning system load collected in step 2, so that the m characteristic elements included in the original data of each air-conditioning system load are all dimensionless data; then, 80% of the original data of the air-conditioning system load collected after de-dimensionalization is used as a training set; and 20% of the original data of the air-conditioning system load collected after de-dimensionalization is used as a test set; The method of de-dimensionalization is: first, each characteristic element is divided into grades, and each grade corresponds to a dimensionless number; then, each acquired characteristic element is matched with the corresponding grade, and the corresponding dimensionless data is obtained; When the characteristic element is season, the binning method is: If it is Spring Festival or Autumn, the dimensionless number is 0; if it is Summer or Winter, the dimensionless number is 1; When the characteristic element is a weekend or holiday, the dimensionless number is 1; otherwise, the dimensionless number is 0; Step 4: Build a deep learning model based on neural network: The deep learning model includes an input layer, y hidden layers and an output layer; the input layer has p neurons, corresponding to p feature elements respectively; p and y are unknown variables to be determined, and p≤m, y≤m; the sigmoid activation function is used as the connection function between layers; the output layer has one neuron, corresponding to the predicted value of the air conditioning system load at the next moment; Step 5, determine the optimal deep learning model: by selecting different p and y values, different deep learning models are obtained; the training set data in step 3 is used to train each deep learning model, and then the connection weight value in each deep learning model is determined; then, the test set data in step 3 is used to verify the model, and different air conditioning system load prediction values ​​are obtained; by comparing or evaluating the obtained different air conditioning system load prediction values, the optimal deep learning model is determined; the number of input layer neurons in the optimal deep learning model is p=p1; The specific method for determining the optimal deep learning model includes the following steps: Step 5A, determine the key characteristic elements: first sort the m characteristic elements selected in step 1 from large to small according to their impact on the load of the air conditioning system; then, take the first p0 characteristic elements as the key characteristic elements; wherein the calculation formula of p0 is: Step 5B, establish a deep learning model: select different p and y values ​​to obtain different deep learning models; where p∈[p0,m], y∈[y0,m], and y0=2; at this time, the p neurons in the input layer of the deep learning model include at least the p0 key feature elements determined in step 5A; Step 5C, calculating the connection weight value: using the training set data in step 3 to train each deep learning model established in step 5B, and updating the connection weight value of each layer based on the gradient descent method, thereby determining the connection weight value in each deep learning model; When updating the connection weight value of each layer based on the gradient descent method, the learning rate α(t) in the gradient descent method is designed according to the distribution characteristics of the Gaussian function. The specific calculation formula is: In the formula, t represents the number of learning iterations, h and q are constants; Step 5D, verifying the deep learning model: using the test set data in step 3 to verify each deep learning model with a determined connection weight value, to obtain different air conditioning system load prediction values; Step 5E, determining the optimal deep learning model: comparing or evaluating the load prediction values ​​of different air-conditioning systems obtained in step 5D, and then obtaining the optimal deep learning model; Step 6, air conditioning system load prediction: obtain p1 characteristic elements of the air conditioning system to be predicted at the current moment and perform dimensionless processing, substitute the dimensionless p1 characteristic elements into the deep learning optimal model determined in step 5, and perform air conditioning system load prediction. The output layer result is the predicted value of the air conditioning system load at the next moment.

2. The air conditioning system load prediction method based on deep learning according to claim 1, characterized in that: In step 4, the expression of the sigmoid activation function is: In the formula, x is the predicted output value of the neuron, and S(x) is the output value of the neuron after nonlinear transformation.

3. The air conditioning system load prediction method based on deep learning according to claim 1, characterized in that: In step 5E, the MAE indicator is used to evaluate the load forecast values ​​of different air-conditioning systems obtained in step 5D. When the MAE is less than 1%, the deep learning model corresponding to it is the optimal deep learning model.

4. The air conditioning system load prediction method based on deep learning according to claim 3 is characterized in that: The calculation formula of MAE is: Where, N is the number of original data of air conditioning system load selected from the test set data in step 5D; i is the i-th original data of air conditioning system load selected from the test set data; Y i is the original measured value of the air conditioning system load at the next moment in the test set data; P i It is the air conditioning system load prediction value obtained by using the deep learning model to predict the i-th air conditioning system load original data in step 5D.

Images