Method for predicting energy consumption of air conditioning system of railway station

Abstract

The invention discloses an energy consumption prediction method for an air conditioning system of a railway station. In the actual operation process of a station building air conditioning system, selection of operation modes of main energy consumption equipment and setting of specific operation parameters have obvious influences on operation energy consumption. According to the method, association mining can be performed on multi-dimensional time sequence data and high-dimensional time sequence features can be extracted by using TCN, meanwhile, the BiGRU model has nonlinear fitting capability, and key information loss can be prevented by performing effective feature extraction on input data. According to the method, more reliable and more accurate energy consumption load look-ahead information is provided for a passenger station energy management system, and high-quality data support is provided for optimizing a control strategy.

Description

Technical Field

[0002] This invention belongs to the field of air conditioning system load prediction technology, specifically relating to a method for predicting the energy consumption of a railway passenger station air conditioning system. Background Technology

[0003] According to data from the International Energy Agency, buildings account for 40% of global energy consumption, and air conditioning systems account for 60% to 75% of the total energy consumption of railway passenger stations. As a major energy-consuming device, the operating efficiency of air conditioning systems has a significant impact on the overall energy consumption of station buildings.

[0004] The air conditioning system of a railway passenger station mainly consists of an air conditioning cold source system, a public area ventilation and air conditioning system, and an equipment and management room air conditioning system. During the actual operation of the station's air conditioning system, the selection of the operating mode and the setting of specific operating parameters (such as supply and return water temperatures, pump frequency, and fan frequency) of the main energy-consuming equipment (chillers, chilled water pumps, cooling towers, cooling water pumps, and terminal units) have a significant impact on operating energy consumption. In recent years, AI algorithms such as neural networks have matured in parameter calculation, and more new algorithms have been developed and used for air conditioning load data prediction, achieving some positive results.

[0005] However, current practical applications are mostly concentrated in the control of air conditioning systems in office buildings, where weekday and non-weekday loads differ from those in railway passenger stations, exhibiting regular variations. Furthermore, due to the lack of universality in various algorithm models, the accuracy of their predictions for large railway passenger station air conditioning systems, which have numerous independent variables and enormous energy consumption, needs improvement. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a method for predicting the energy consumption of air conditioning systems in railway passenger stations. By using TCN, multidimensional time series data can be correlated and mined to extract high-dimensional time series features. At the same time, the BiGRU model has nonlinear fitting capabilities. By effectively extracting features from the input data, key information loss can be prevented and prediction accuracy can be improved.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting the energy consumption of a railway passenger station air conditioning system includes the following steps: Step 1: Obtain historical data on power consumption or efficiency of each air conditioning unit in the passenger station, as well as indoor and outdoor temperature and humidity, wind speed, solar radiation intensity, and real-time passenger flow. Step 2: Clean and normalize the data collected in Step 1; Step 3: Calculate the Pearson correlation coefficient to assess the correlation between each feature and the target load value; select features with high correlation to the target load to form the final feature vector; select feature elements that affect the air conditioning system load: through analysis of the air conditioning system to be predicted, determine m feature elements that affect the air conditioning system load; Step 4: Dataset construction. Input sequence length and prediction step size, and divide the dataset into training set, validation set and test set in a 7:2:1 ratio. Step 5: Establish a TCN-BiGRU prediction model. After reconstructing the data, feature extraction is performed by TCN and weight normalization is performed. The ReLU unit function is used as the activation function. The features extracted by TCN are input into BiGRU through a fully connected layer for learning to extract the internal time variation pattern. Step Six: Establish an evaluation index system to comprehensively judge the model's predictive performance; Step 7: Input the processed sequence into the model for prediction and evaluation.

[0008] Furthermore, step two specifically involves: using the 3σ criterion to remove outlier data, using linear interpolation to fill in missing values, and using linear normalization to transform the input data to ensure data consistency.

[0009] Furthermore, in step three, the Pearson correlation coefficient is calculated to evaluate the correlation between each feature and the target load value, using the following formula: In the formula: X i and Y i They are two variables. X and Y The i-th observation; and They are X and Y The sample mean; n is the number of samples; m=7, and the 7 characteristic elements are historical load value, temperature, humidity, wind speed, season, solar irradiance, and real-time passenger flow.

[0010] Furthermore, in step four, the sequence length and prediction step size are defined; the sample structure is: X=[feature(tn), ..., feature(t-1), feature(t)], y=[load(t+1), ..., load(t+h)].

[0011] Furthermore, in step five, the TCN includes causal convolution, dilated convolution, and a residual module. First, causal convolution is used to ensure the temporal nature of the convolution result, and dilation coefficients are introduced to adjust it to adapt to different time series. One-dimensional data x is fed into the causal convolution and trained using a one-dimensional convolution kernel. X t To predict Y t The expression is as follows: In the formula, filter F=( f 1, ..., f k ); Data sequence X = ( x 1, ..., x t ),exist x t The dilated convolution operation is: In the formula, This indicates convolution over past states; K is the filter size; k is the number of network layers; d is the dilation coefficient. A residual block with input x and output y is defined as follows: In the formula, H Represents a series of convolution operation functions; W The weights represent the weights of the convolution kernel. W The formula for parameter rewriting is: In the formula, h It is a scalar; Let x be the Euclidean norm; W Both x and y are k-dimensional vectors.

[0012] Furthermore, in step five, BiGRU is a GRU network architecture with a dual structure, responsible for forward and reverse information respectively; In the forward GRU layer, the input time series is fed into the network for training via forward propagation; in the backward GRU layer, the input sequence is trained via backward propagation.

[0013] Furthermore, in step six, three evaluation indicators are used to measure the degree of agreement between the predicted value and the actual value: root mean square error, mean absolute percentage error, and coefficient of determination.

[0014] Furthermore, in step two, the data cleaning and normalization formulas are as follows: 1) 3σ criterion In the formula: m is the number of samples; X k For sample data; σ is the mean; σ is the standard deviation. 2) Linear interpolation is used to fill in missing values. Linear normalization is used to transform the input data, as shown in the following formula: In the formula: X n Represents the nth data point; X i These are the original data values ​​before linear normalization; These are data values ​​after linear normalization; X max , X min These are the maximum and minimum values ​​in the current parameters, respectively.

[0015] The beneficial effects of this invention are: 1) The TCN-BiGRU model in this invention combines the functions of a "magnifying glass" (TCN excels at spatial and local temporal features) and a "wide-angle lens" (BiGRU excels at global temporal dependencies), enabling it to comprehensively and meticulously capture the spatiotemporal factors affecting energy consumption in the railway passenger station environment, laying a solid foundation for accurate prediction. 2) This invention can not only accurately predict regular daily cycle fluctuations, but also better predict abnormal energy consumption peaks caused by special events (such as holidays and large-scale events), making the prediction results more consistent with the actual operating conditions, providing more reliable and accurate forward-looking information on energy consumption load for the passenger station energy management system, and providing high-quality data support for optimizing control strategies. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the operation of the present invention. Figure 2 This is a diagram of the model architecture of the present invention; Figure 3 This is a structural diagram of a single BiGRU neuron. Detailed Implementation

[0017] The present invention will now be described in detail with reference to specific embodiments.

[0018] This invention provides a method for predicting the energy consumption of air conditioning systems in railway passenger stations. By using TCN, correlation mining can be performed on multidimensional time series data to extract high-dimensional time series features. At the same time, the BiGRU model has nonlinear fitting capabilities. By effectively extracting features from the input data, key information loss can be prevented and prediction accuracy can be improved.

[0019] like Figure 1 As shown, the energy consumption prediction method for railway passenger station air conditioning systems of the present invention includes the following steps: Step 1: Use sensors or data acquisition systems to obtain historical data on the power consumption or cooling / heating power of each air conditioning unit in the passenger station, as well as indoor and outdoor temperature and humidity, wind speed, solar radiation intensity, and real-time passenger flow. Step Two: Clean and normalize the data collected in Step One; specifically: use the 3σ criterion to remove outliers, use linear interpolation to fill in missing values, and use linear normalization to transform the input data to ensure data consistency; the data cleaning and normalization formulas are as follows: 1) 3σ criterion (Laida criterion) In the formula: m is the number of samples; X k For sample data; σ is the mean; σ is the standard deviation.

[0020] 2) Linear interpolation is used to fill in missing values. Linear normalization is used to transform the input data, as shown in the following formula: In the formula: X n Represents the nth data point; X i These are the original data values ​​before linear normalization; These are data values ​​after linear normalization; X max , X min These are the maximum and minimum values ​​in the current parameters, respectively.

[0021] Step 3: Calculate the Pearson correlation coefficient to assess the correlation between each feature and the target load value; select features with high correlation to the target load to form the final feature vector; select feature elements affecting the air conditioning system load: through analysis of the air conditioning system to be predicted, determine m feature elements affecting the air conditioning system load; the formula is as follows: In the formula: X i and Y i They are two variables. X and Y The i-th observation; and They are X and Y The sample mean; n is the number of samples; r The correlation coefficient ranges from [-1, 1]. r =1: indicates a perfect positive linear correlation, meaning there is a perfect positive linear relationship between the two variables; r =-1: indicates a completely negative linear correlation, meaning there is a completely negative linear relationship between the two variables; r =0: indicates that there is no linear relationship between the two variables; m=7, and the 7 characteristic elements are historical load value, temperature, humidity, wind speed, season, solar irradiance, and real-time passenger flow.

[0022] Step 4: Dataset construction. Input sequence length (L) and prediction step size (H). Divide the dataset into training set, validation set and test set in a 7:2:1 ratio. Define sequence length (L): historical time step 72 hours, prediction step size (H): future prediction duration 24 hours. The sample structure is: X=[feature(tn), ..., feature(t-1), feature(t)], y=[load(t+1), ..., load(t+h)].

[0023] Step 5: Establish a TCN-BiGRU prediction model. After reconstructing the data, feature extraction is performed through TCN, and weight normalization is applied. The ReLU unit function is used as the activation function. The features extracted by TCN are input into BiGRU through a fully connected layer for learning, extracting internal time variation patterns. By using TCN to perform correlation mining on multidimensional time series data, high-dimensional time series features are extracted. As a variant of GRU, the BiGRU model has non-linear fitting capabilities. By effectively extracting features from the input data, key information loss can be prevented, and the ability to mine time series relationships can be enhanced.

[0024] 1) TCN includes causal convolution, dilated convolution, and residual modules. First, causal convolution is used to ensure the temporal nature of the convolution results. Dilation coefficients are introduced to adjust its adaptability to different time series. The residual module is capable of preserving the core information characteristics of the time series, thereby extending the effective input time of the time series, reducing computational complexity, and effectively solving the problems of gradient explosion and gradient vanishing caused by shared parameters across different time periods. One-dimensional data x is fed into the causal convolution and trained using a one-dimensional convolution kernel. X t To predict Y t The expression is as follows: In the formula, filter F=( f 1, ..., f k ); Data sequence X = ( x 1, ..., x t ),exist x t The dilated convolution operation is: In the formula, This represents convolution over past states; K is the filter size; k is the number of network layers; d is the dilation coefficient; residual connections ensure the accuracy of deep networks by coupling the input and output information of the convolution operation; a residual block with input x and output y is defined as follows: In the formula, H Represents a series of convolution operation functions; W The weights represent the weights of the convolution kernel. W The formula for parameter rewriting is: In the formula, h It is a scalar; Let x be the Euclidean norm; W Both x and y are k-dimensional vectors.

[0025] Each residual block contains two layers of causal convolutional models. A spatial dropout layer is added inside the residual block to randomly drop a portion of neurons according to a set dropout rate, which can reduce redundant features and avoid overfitting.

[0026] 2) BiGRU is a GRU network architecture with a dual structure, responsible for forward and reverse information respectively; see the BiGRU network structure diagram. Figure 2 The middle BiGRU layer is shown in the figure; X tt represents the input data at time t; n is the total length of the sequence; This represents the output of the forward hidden layer at time t; This represents the output of the reverse hidden layer at time t; O The output of BiGRU is composed of the forward hidden layer at the last time step and the backward hidden layer at the first time step.

[0027] In the forward GRU layer, the input time series is fed into the network for training through forward propagation, with the aim of acquiring information about the sequence in the past. In the backward GRU layer, the input sequence is trained through backward propagation, with the aim of acquiring information about the sequence's potential future occurrence. Since the backward propagation process can process input signals in two directions simultaneously, the network structure can extract bidirectional temporal features from the input data, thereby improving the comprehensiveness and completeness of feature extraction.

[0028] The structure of a single BiGRU neuron is as follows Figure 3 As shown, the definition and interpretation of a single BiGRU are as follows: In the formula: σ is the sigmoid activation function, used to restrict the gate value to between 0 and 1, and the tanh activation function is used for updating the hidden state. r t and u t For the reset gate and update gate at time t: H t and H t-1 It represents the hidden state at time t and the previous time. W It is the input weight matrix, used to connect the input data and the hidden state of the previous time step, and used to integrate features during the calculation process; U As a shared weight matrix, it is multiplied only with the hidden state of the previous time step and used to calculate the gating information for the forward and backward directions; b It is a bias term.

[0029] Step Six: Establish an evaluation index system to comprehensively judge the model's predictive performance; use root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 Three evaluation indicators are used to measure the degree of agreement between predicted and actual values, as follows: In the formula: n represents the number of prediction points; The actual value; This is a predicted value; This represents the average value.

[0030] Step 7: Input the processed sequence into the model for prediction and evaluation.

[0031] The content of this invention is not limited to the embodiments listed. Any equivalent modifications made by those skilled in the art to the technical solutions of this invention by reading this specification are covered by the claims of this invention.

Claims

1. A method for predicting energy consumption of a railway passenger station air conditioning system, characterized in that: It includes the following steps: Step 1: Obtain historical data on power consumption or efficiency of each air conditioning unit in the passenger station, as well as indoor and outdoor temperature and humidity, wind speed, solar radiation intensity, and real-time passenger flow. Step 2: Clean and normalize the data collected in Step 1; Step 3: Calculate the Pearson correlation coefficient to assess the correlation between each feature and the target load value; select features with high correlation to the target load to form the final feature vector; select feature elements that affect the air conditioning system load: through analysis of the air conditioning system to be predicted, determine m feature elements that affect the air conditioning system load; Step 4: Dataset construction. Input sequence length and prediction step size, and divide the dataset into training set, validation set and test set in a 7:2:1 ratio. Step 5: Establish a TCN-BiGRU prediction model. After reconstructing the data, feature extraction is performed by TCN and weight normalization is performed. The ReLU unit function is used as the activation function. The features extracted by TCN are input into BiGRU through a fully connected layer for learning to extract the internal time variation pattern. Step Six: Establish an evaluation index system to comprehensively judge the model's predictive performance; Step 7: Input the processed sequence into the model for prediction and evaluation.

2. The method for predicting energy consumption of a railway passenger station air conditioning system according to claim 1, characterized in that: Step two specifically involves: using the 3σ criterion to remove outlier data, using linear interpolation to fill in missing values, and using linear normalization to transform the input data to ensure data consistency.

3. The method for predicting energy consumption of a railway passenger station air conditioning system according to claim 2, characterized in that: In step three, the Pearson correlation coefficient is calculated to assess the correlation between each feature and the target load value. The formula is as follows: In the formula: X i and Y i They are two variables. X and Y The i-th observation; and They are X and Y The sample mean; n is the number of samples; m=7, and the 7 characteristic elements are historical load value, temperature, humidity, wind speed, season, solar irradiance, and real-time passenger flow.

4. The method for predicting energy consumption of a railway passenger station air conditioning system according to claim 3, characterized in that: In step four, the sequence length and prediction step size are defined; the sample structure is: X=[feature(tn), ..., feature(t-1), feature(t)], y=[load(t+1), ..., load(t+h)].

5. The method for predicting energy consumption of a railway passenger station air conditioning system according to claim 4, characterized in that: In step five, the TCN includes causal convolution, dilated convolution, and residual modules. First, causal convolution is used to ensure the temporal nature of the convolution results. Dilation coefficients are introduced to adjust its adaptability to different time series. One-dimensional data x is fed into the causal convolution and trained using a one-dimensional convolution kernel. X t To predict Y t The expression is as follows: In the formula, filter F=( f 1, ..., f k ); Data sequence X = ( x 1, ..., x t ),exist x t The dilated convolution operation is: In the formula, This indicates convolution over past states; K is the filter size; k is the number of network layers; d is the dilation coefficient. A residual block with input x and output y is defined as follows: In the formula, H Represents a series of convolution operation functions; W The weights represent the weights of the convolution kernel. W The formula for parameter rewriting is: In the formula, h It is a scalar; Let x be the Euclidean norm; W Both x and y are k-dimensional vectors.

6. The method for predicting energy consumption of a railway passenger station air conditioning system according to claim 5, characterized in that: In step five, BiGRU is a GRU network architecture with a dual structure, responsible for forward and reverse information respectively. In the forward GRU layer, the input time series is fed into the network for training via forward propagation; in the backward GRU layer, the input sequence is trained via backward propagation.

7. The method for predicting energy consumption of a railway passenger station air conditioning system according to claim 6, characterized in that: In step six, three evaluation indicators are used to measure the degree of agreement between the predicted value and the actual value: root mean square error, mean absolute percentage error, and coefficient of determination.

8. The method for predicting energy consumption of a railway passenger station air conditioning system according to claim 7, characterized in that: In step two, the data cleaning and normalization formulas are as follows: 1) 3σ criterion In the formula: m is the number of samples; X k For sample data; σ is the mean; σ is the standard deviation. 2) Linear interpolation is used to fill in missing values. Linear normalization is used to transform the input data, as shown in the following formula: In the formula: X n Represents the nth data point; X i These are the original data values ​​before linear normalization; These are data values ​​after linear normalization; X max , X min These are the maximum and minimum values ​​in the current parameters, respectively.

Images