Enterprise atmospheric pollutant emission prediction method and system based on power load data

Abstract

The invention provides an enterprise atmospheric pollutant emission prediction method and system based on power load data. The method comprises the following steps: acquiring current power load data of an enterprise; inputting the current power load data into a short-term power prediction model for load prediction to obtain current short-term power prediction data; inputting the current short-term power prediction data into an atmospheric pollutant emission prediction model for prediction to obtain an atmospheric pollutant emission prediction result of the enterprise; the short-term power prediction model is composed of a variational mode decomposition (VMD) algorithm, a convolutional neural network (CNN) layer and a bidirectional long short-term memory (BiLSTM) layer; the atmospheric pollutant emission prediction model is constructed based on a particle swarm optimization PSO algorithm and a support vector machine regression SVR model. The short-term electric power prediction model and the atmospheric pollutant emission prediction model do not need a large amount of data, so that the consumption of computing resources can be reduced, the generalization capability is relatively high, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of air pollutant prevention and control, and in particular to a method and system for predicting enterprise air pollutant emissions based on power load data. Background Art

[0002] As the core production factor of an enterprise, the consumption data of electric energy can reflect the production activities of the enterprise. With the development of smart grid and information construction, the power industry has accumulated a large amount of structured, semi-structured and unstructured business data, which has the characteristics of big data in terms of data volume, diversity, speed and value. Electric power big data covers all aspects such as power generation, transmission, distribution and power consumption, including power generation, power consumption, equipment status, environmental parameters, etc. The development of big data technology has shown new application prospects in power system operation optimization, pollution source identification, environmental supervision and other aspects. The use of advanced technical means such as big data analysis, artificial intelligence, computer science, etc. to predict and control the emission of atmospheric pollutants from enterprises is of great significance to protecting the environment and achieving sustainable development.

[0003] At present, historical pollutant emission data is usually used to establish a regression prediction model represented by statistics and machine learning, and then the established regression prediction model is used for prediction. The establishment of such models requires a large amount of data, and the large amount of data processing will lead to a large consumption of computing resources. In addition, the generalization ability of such models is limited. If the production process, equipment status or environmental conditions of the enterprise change, the prediction accuracy will be reduced. Summary of the invention

[0004] In order to overcome the defects of large amount of calculation, high consumption of computing resources, limited generalization ability and low prediction accuracy required for the above-mentioned regression prediction models currently represented by statistics and machine learning when predicting pollutant emissions, the present invention provides a method for predicting enterprise atmospheric pollutant emissions based on power load data, the method comprising:

[0005] Obtain the current power load data of the enterprise;

[0006] Inputting the current power load data into a short-term power forecasting model to perform load forecasting to obtain current short-term power forecasting data;

[0007] Inputting the current short-term power forecast data into an air pollutant emission forecast model for forecasting to obtain an air pollutant emission forecast result for the enterprise;

[0008] The short-term power forecasting model is composed of a variational mode decomposition (VMD) algorithm, a convolutional neural network (CNN) layer, and a bidirectional long short-term memory (BiLSTM) network layer; the atmospheric pollutant emission forecasting model is constructed based on a particle swarm optimization (PSO) algorithm and a support vector machine regression (SVR) model.

[0009] Optionally, the process of constructing the short-term power forecasting model includes:

[0010] Obtain historical power load data sets;

[0011] Using a time series construction method, historical power load data of a specific duration and future short-term power data corresponding to the historical power load data are selected from the historical power load data set;

[0012] The VMD algorithm in the initial short-term power forecasting model is used to perform signal decomposition on the historical power load data to obtain multiple modal components of the historical power data;

[0013] Using the one-dimensional CNN layer in the initial short-term power forecasting model to extract features of the multiple historical power data modal components to obtain historical power load data features;

[0014] Using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain future short-term power forecasting data;

[0015] Based on the future short-term power data and the future short-term prediction data, the initial short-term power prediction model is iteratively trained to obtain a final short-term power prediction model.

[0016] Optionally, the time series construction method is used to select historical power load data of a specific duration and future short-term power data corresponding to the historical power load data from the historical power load data set, including:

[0017] Normalizing the historical power load data set using a maximum-minimum normalization algorithm to obtain a normalized historical power load data set;

[0018] Using a time series construction method to convert the normalized historical power load data set into a supervised learning sequence;

[0019] In the supervised learning sequence, historical power load data of a specific duration and future short-term power data corresponding to the historical power load data are selected.

[0020] Optionally, the using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain future short-term power forecasting data includes:

[0021] Using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain multiple load forecasting results;

[0022] The multiple load forecast results are summed to obtain future short-term power forecast data.

[0023] Optionally, after iteratively training the initial short-term power prediction model based on the future short-term power data and the future short-term prediction data to obtain a final short-term power prediction model, the method further includes:

[0024] Obtain historical power load forecast sets;

[0025] Inputting the historical power load forecast set into the short-term power forecast model to perform load forecasting, and obtaining a power forecast data sequence of the historical power load forecast set;

[0026] Performing a denormalization process on the power prediction data sequence to obtain a denormalized power prediction data sequence;

[0027] The short-term power forecasting model is evaluated based on the real power data sequence in the historical power load forecasting set and the denormalized power forecasting data sequence.

[0028] Optionally, after the maximum-minimum normalization algorithm is used to normalize the historical power load data set to obtain the normalized historical power load data set, the method further includes:

[0029] Analyzing the characteristic correlation of the characteristic parameters in the normalized historical power load data set using the Pearson coefficient analysis method to obtain the correlation coefficient of each characteristic parameter;

[0030] Based on each characteristic parameter, if the correlation coefficient of the characteristic coefficient exceeds the correlation coefficient threshold, the historical power load data of the characteristic parameter is saved in the historical power load data set, otherwise the historical power load data of the characteristic parameter is deleted from the historical power load data set.

[0031] Optionally, the method for constructing the atmospheric pollutant emission prediction model includes:

[0032] Obtaining a pollutant emission labeling data set, wherein the pollutant emission labeling data set includes historical power load data and pollutant emission monitoring data of the same period;

[0033] The PSO algorithm is used to optimize the initial hyperparameters of the SVR model to obtain the optimal hyperparameters;

[0034] The historical power load data and the pollutant emission monitoring data of the same period are used to iteratively train the SVR model with the optimal hyperparameters to obtain an atmospheric pollutant emission prediction model.

[0035] Optionally, the PSO algorithm is used to optimize the initial hyperparameters of the SVR model to obtain the optimal hyperparameters, including:

[0036] Initialize the position and velocity of each particle in the PSO algorithm. Each particle represents a set of initial hyperparameters of the SVR model.

[0037] The fitness value of each particle's position is calculated according to the fitness function, and the particle's speed and position are updated to obtain a new individual optimal position and a global optimal position of the particle group. The particle's speed and position are updated repeatedly until the iteration termination condition is reached. The particle with the best fitness value when the iteration termination condition is reached is used as the optimal hyperparameter of the SVR model.

[0038] Optionally, the iteration termination condition includes obtaining a maximum number of iterations, or the optimal fitness does not change in a set number of consecutive iterations.

[0039] On the other hand, the present invention also provides an enterprise atmospheric pollutant emission prediction system based on power load data, comprising:

[0040] An acquisition module is used to obtain the current power load data of the enterprise;

[0041] A load forecasting module, used for inputting the current power load data into a short-term power forecasting model to perform load forecasting and obtain current short-term power forecasting data;

[0042] A pollutant emission prediction module, used to input the current short-term power prediction data into the atmospheric pollutant emission prediction model for prediction, and obtain the atmospheric pollutant emission prediction result of the enterprise;

[0043] The short-term power forecasting model is composed of a variational mode decomposition (VMD) algorithm, a convolutional neural network (CNN) layer, and a bidirectional long short-term memory (BiLSTM) network layer; the atmospheric pollutant emission forecasting model is constructed based on a particle swarm optimization (PSO) algorithm and a support vector machine regression (SVR) model.

[0044] Optionally, it also includes a short-term power forecast model building module for:

[0045] Obtain historical power load data sets;

[0046] Using a time series construction method, historical power load data of a specific duration and future short-term power data corresponding to the historical power load data are selected from the historical power load data set;

[0047] The VMD algorithm in the initial short-term power forecasting model is used to perform signal decomposition on the historical power load data to obtain multiple modal components of the historical power data;

[0048] Using the one-dimensional CNN layer in the initial short-term power forecasting model to extract features of the multiple historical power data modal components to obtain historical power load data features;

[0049] Using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain future short-term power forecasting data;

[0050] Based on the future short-term power data and the future short-term prediction data, the initial short-term power prediction model is iteratively trained to obtain a final short-term power prediction model.

[0051] Optional short-term power forecast model building module, specifically for:

[0052] Normalizing the historical power load data set using a maximum-minimum normalization algorithm to obtain a normalized historical power load data set;

[0053] Using a time series construction method to convert the normalized historical power load data set into a supervised learning sequence;

[0054] In the supervised learning sequence, historical power load data of a specific duration and future short-term power data corresponding to the historical power load data are selected.

[0055] Optional short-term power forecast model building module, specifically for:

[0056] Using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain multiple load forecasting results;

[0057] The multiple load forecast results are summed to obtain future short-term power forecast data.

[0058] Optional, short-term power forecasting model building module is also used to:

[0059] Obtain historical power load forecast sets;

[0060] Inputting the historical power load forecast set into the short-term power forecast model to perform load forecasting, and obtaining a power forecast data sequence of the historical power load forecast set;

[0061] Performing a denormalization process on the power prediction data sequence to obtain a denormalized power prediction data sequence;

[0062] The short-term power forecasting model is evaluated based on the real power data sequence in the historical power load forecasting set and the denormalized power forecasting data sequence.

[0063] Optional, short-term power forecasting model building module is also used to:

[0064] Analyzing the characteristic correlation of the characteristic parameters in the normalized historical power load data set using the Pearson coefficient analysis method to obtain the correlation coefficient of each characteristic parameter;

[0065] Based on each characteristic parameter, if the correlation coefficient of the characteristic coefficient exceeds the correlation coefficient threshold, the historical power load data of the characteristic parameter is saved in the historical power load data set, otherwise the historical power load data of the characteristic parameter is deleted from the historical power load data set.

[0066] Optionally, it also includes an atmospheric pollutant emission prediction model building module for:

[0067] Obtaining a pollutant emission labeling data set, wherein the pollutant emission labeling data set includes historical power load data and pollutant emission monitoring data of the same period;

[0068] The PSO algorithm is used to optimize the initial hyperparameters of the SVR model to obtain the optimal hyperparameters;

[0069] The historical power load data and the pollutant emission monitoring data of the same period are used to iteratively train the SVR model with the optimal hyperparameters to obtain an atmospheric pollutant emission prediction model.

[0070] Optional, atmospheric pollutant emission prediction model building module, specifically used for:

[0071] Initialize the position and velocity of each particle in the PSO algorithm. Each particle represents a set of initial hyperparameters of the SVR model.

[0072] The fitness value of each particle's position is calculated according to the fitness function, and the particle's speed and position are updated to obtain a new individual optimal position and a global optimal position of the particle group. The particle's speed and position are updated repeatedly until the iteration termination condition is reached. The particle with the best fitness value when the iteration termination condition is reached is used as the optimal hyperparameter of the SVR model.

[0073] Optionally, the iteration termination condition includes obtaining a maximum number of iterations, or the optimal fitness does not change in a set number of consecutive iterations.

[0074] On the other hand, the present invention also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus;

[0075] The memory is used to store one or more programs;

[0076] When the one or more programs are executed by the at least one processor, any one of the above-mentioned methods for predicting enterprise atmospheric pollutant emissions based on power load data is implemented.

[0077] On the other hand, the present invention also provides a readable storage medium having an execution program stored thereon, and when the execution program is executed, any of the above-mentioned methods for predicting enterprise atmospheric pollutant emissions based on power load data is implemented.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] The present invention provides a method and system for predicting enterprise atmospheric pollutant emissions based on power load data, including: obtaining the current power load data of the enterprise; inputting the current power load data into a short-term power forecasting model for load forecasting to obtain the current short-term power forecasting data; inputting the current short-term power forecasting data into an atmospheric pollutant emission forecasting model for forecasting to obtain the atmospheric pollutant emission forecasting result of the enterprise; the short-term power forecasting model is composed of a variational mode decomposition VMD algorithm, a convolutional neural network CNN layer and a bidirectional long short-term memory network BiLSTM layer; the atmospheric pollutant emission forecasting model is constructed based on a particle swarm optimization PSO algorithm and a support vector machine regression SVR model. The short-term power forecasting model of the present invention processes the high volatility and nonlinearity of power load data through VMD decomposition, the BiLSTM layer can predict the intrinsic relationship of the time series in the past and future two dimensions, and the SVR in the atmospheric pollutant emission forecasting model is suitable for the prediction of a small number of samples, so the short-term power forecasting model and the atmospheric pollutant emission forecasting model do not require a large amount of data, so the consumption of computing resources can be reduced and the generalization ability is high, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It is a flow chart of the method for predicting enterprise atmospheric pollutant emissions based on power load data of the present invention;

[0081] Figure 2 It is a schematic diagram of VMD decomposition results of the present invention;

[0082] Figure 3 It is a schematic diagram of the IMF component spectrum of the present invention;

[0083] Figure 4 Schematic diagram of the working principle of 1D-CNN of the present invention;

[0084] Figure 5 It is a basic structural diagram of the LSTM layer of the present invention;

[0085] Figure 6 It is a schematic diagram of the structure of the BiLSTM of the present invention;

[0086] Figure 7 It is a flow chart of the VMD-CNN-BiLSTM model of the present invention;

[0087] Figure 8 It is a schematic diagram of the process of the method for predicting the emission of atmospheric pollutants from enterprises of the present invention;

[0088] Fig. 9 It is a schematic diagram of the sulfur dioxide emission prediction result of the present invention;

[0089] Fig.10 It is a schematic diagram of the nitrogen oxide emission prediction result of the present invention;

[0090] Fig.11 It is a schematic diagram of smoke emission prediction results of the present invention;

[0091] Fig.12 It is a structural schematic diagram of the enterprise atmospheric pollutant emission prediction system based on power load data of the present invention;

[0092] Fig.13 It is a schematic structural diagram of the electronic device of the present invention. DETAILED DESCRIPTION

[0093] The specific implementation modes of the present invention are further described in detail below with reference to the accompanying drawings.

[0094] Embodiment 1:

[0095] The present invention provides a method for predicting enterprise atmospheric pollutant emissions based on power load data, the flow chart of which is as follows: Figure 1 As shown, including:

[0096] Step 101: Obtain the current power load data of the enterprise.

[0097] Step 102: Input the current power load data into the short-term power forecasting model to perform load forecasting, and obtain the current short-term power forecast data of the enterprise.

[0098] Step 103: Input the current short-term power forecast data into the atmospheric pollutant emission forecast model for forecasting, and obtain the forecast results of the atmospheric pollutant emission of the enterprise. The short-term power forecast model is composed of the variational mode decomposition (VMD) algorithm, the convolutional neural network (CNN) layer and the bidirectional long short-term memory network (BiLSTM) layer; the atmospheric pollutant emission forecast model is constructed based on the particle swarm optimization (PSO) algorithm and the support vector regression (SVR) model.

[0099] The production behavior of an enterprise is directly related to its pollutant emissions, and electricity is the energy source of the enterprise's production. In other words, pollutants are the products of the enterprise's production process. Modern industrial enterprises will inevitably consume a lot of electricity in the production process. Therefore, based on the relationship between the enterprise's electricity consumption and the emission of atmospheric pollutants, the enterprise's atmospheric pollution emissions can be accurately predicted by predicting the enterprise's electricity consumption. In addition, the short-term power forecasting model of the embodiment of the present invention processes the high volatility and nonlinearity of power load data through VMD decomposition. The BiLSTM layer can predict the intrinsic relationship between the time series in the past and future dimensions. The SVR in the atmospheric pollutant emission prediction model is suitable for the prediction of a small number of samples. Therefore, the short-term power forecasting model and the atmospheric pollutant emission prediction model do not require a large amount of data, so it can reduce the consumption of computing resources and have a higher generalization ability, thereby improving the prediction accuracy.

[0100] The current power load data of the enterprise in step 101 may be the power consumption of power users in a certain area within a set time range. For example, but not limited to, the power load data may be collected from the State Grid Information Center and from the user's smart meter, with a collection frequency of once every 15 minutes. In the embodiment of the present invention, the power users may cover 1,000 (only for example, not limited in actual implementation) key industrial enterprises in the area.

[0101] In a possible implementation, the process of constructing the short-term power forecasting model in S102 includes:

[0102] Obtain historical power load data sets;

[0103] The time series construction method is used to select historical power load data of a specific time period and future short-term power data corresponding to the historical power load data from the historical power load data set;

[0104] The VMD algorithm in the initial short-term power forecasting model is used to decompose the historical power load data to obtain multiple historical power data modal components (Intrinsic Mode Function, IMF);

[0105] The one-dimensional CNN (1D-CNN) layer in the initial short-term power forecasting model is used to extract features from multiple historical power data modal components to obtain the features of historical power load data;

[0106] The BiLSTM layer in the initial short-term power forecasting model is used to forecast the historical power load data characteristics to obtain future short-term power forecast data;

[0107] Based on future short-term power data and future short-term forecast data, the initial short-term power forecast model is iteratively trained to obtain the final short-term power forecast model.

[0108] In this implementation process, the short-term power forecasting model built is the VMD-CNN-BILSTM power data short-term forecasting model. The purpose is to ① use the VMD decomposition module to deal with problems such as high volatility and nonlinearity of power load data; ② use the BiLSTM network to perform actual forecasting of power (load) data, so that the model can learn the intrinsic relationship of data from the past and future dimensions of the time series in both directions, effectively improve the model's forecasting ability and data utilization, and improve the model's generalization ability while ensuring the model's forecasting accuracy. In an example, the CNN layer can be trained first and then the BiLSTM layer can be trained to complete the training of the initial short-term power forecasting model.

[0109] In the implementation process, when obtaining the historical power load data set, it can be the historical power data obtained, for example, including but not limited to: randomly selecting the 24-hour power consumption of a power user in a certain city throughout the year (this study considers that the hourly power consumption is the hourly load) as the power load data set.

[0110] The time series construction method can convert the historical power load data set into a supervised learning sequence, and then select a specific input duration and corresponding data features, and then perform VMD decomposition on the input duration data as input data to obtain several power data modal components IMF. Optionally, in an embodiment of the present invention, a training set and a test set can be divided for the acquired historical power load data set, or a training set and a test set can be constructed for each power data modal component, the training set is used for iterative training of the short-term power forecasting model, and the test set is used for evaluating the short-term power forecasting model.

[0111] Considering that different data features often have different dimensions, the difference between data may be large. In order to eliminate the influence of the difference in dimension and value range, the collected historical power load data set can be normalized, and then the time series construction method can be used to obtain the characteristic data set about the power load for model training. For example, when the time series construction method is used in the historical power load data set, the historical power load data set of a specific length and the future short-term power data corresponding to the historical power load data are selected, the maximum and minimum normalization algorithm can be used to normalize the historical power load data set to obtain the normalized historical power load data set; the time series construction method is used to convert the normalized historical power load data set into a supervised learning sequence; in the supervised learning sequence, the historical power load data of a specific length and the future short-term power data corresponding to the historical power load data are selected. This process can be regarded as a data preprocessing process. After the data is normalized, the time series construction method is constructed. First, a time length window is selected, and the historical data of a specific length before the specific load data point is selected by sliding. This part of the data is the feature input of the model, and the load data point is the model label. Specifically, in the supervised learning sequence, a time length window can be selected, and historical data of a specific length before a specific load data point can be slid and selected. This part of the data is the feature input of the model, and the load data point is the model label. By selecting the relatively optimal input time window, the accuracy of the model can be further optimized. After comparison, the model's relatively optimal input time window length is 48 hours (the length gradient is 24 hours), and the model error evaluation index is the smallest at this time. This optimal time window length may not be unique for different objects. For example, industrial users will be directly related to their production laws, while home users will be related to their life laws. When targeting different users, the window length can be adjusted object-oriented.

[0112] In an optional implementation, after the maximum and minimum standardization algorithm is used to normalize the historical power load data set to obtain the normalized historical power load data set, the Pearson coefficient analysis method can also be used to analyze the characteristic correlation of the characteristic parameters in the normalized historical power load data set to obtain the correlation coefficient of each characteristic parameter; based on each characteristic parameter, if the correlation coefficient of the characteristic coefficient exceeds the correlation coefficient threshold, the historical power load data of the characteristic parameter is saved in the historical power load data set, otherwise the historical power load data of the characteristic parameter is deleted from the historical power load data set. The characteristic parameters may include historical load, population growth rate, socio-economic situation, geographical environment, temperature at the time of prediction, historical rainfall, wind speed, relative humidity, month, week, date, holiday, time, etc. Taking the characteristic parameter of historical load as an example, historical load is important information in the load forecasting model. It is not only the direct input data of the model, but also directly reflects the possible changes in the load in the future. After analysis, the characteristic correlation of historical load decreases with the increase of time interval, but its correlation coefficient r is above 0.92, indicating that the characteristics of each historical load are significant characteristics. For other characteristic parameters, the correlation coefficient r can also be obtained through correlation analysis, and it is selected to be added or not added to the input feature. The historical power load data of the characteristic parameters added to the input feature are retained in the historical power load data set, and the historical power load data of the characteristic parameters not added to the input feature can be deleted in the historical power load data set. It can be understood that the specific value of the correlation coefficient threshold is not limited here.

[0113] The historical power load data of a specific time length selected by the time series construction method can be regarded as the feature number input of the model, and the future short-term power data corresponding to the historical power load data can be used as the model label, so that the model can learn the intrinsic relationship of the data in both the past and future dimensions of the time series.

[0114] In the existing time series prediction methods, although a variety of deep neural network models can be used for load prediction, the decomposition prediction method is currently mostly used for time series prediction based on single user variables. Observing the original sequence, it can be seen that the load data does not have strong periodicity and the data fluctuation is complex. At the same time, the data will contain a certain degree of noise, which brings difficulties to the load prediction work. In the embodiment of the present invention, the problem of spectrum aliasing can be avoided by performing VMD decomposition on the historical power load data. In addition, VMD has the advantage of being able to independently select the number of modes.

[0115] For example, the VMD decomposition process can be shown in the following formula:

[0116] In the first step, the selected historical power load data is subjected to the Hilbert transform shown in the following formula (1) to obtain the analyzed data:

[0117]

[0118] In the formula, * is the convolution operation; j represents the imaginary unit; t is the time variable; δ(t) is the unit pulse signal; u k 、u k (t) are the decomposed k mode functions belonging to the historical power load data f.

[0119] In the second step, the following formula (2) is used to calculate the modal functions u in the previous step: k The center frequency of (t) is estimated and the spectrum is modulated on the corresponding modal components:

[0120]

[0121] In the formula, ω k ={ω k |ω1,…,ω k} is the center frequency of each modal component. Based on formula (2), the variational problem is constructed, and then the bandwidth of each modal component is calculated using the following formula (3):

[0122]

[0123] In the formula, Denotes the derivative operation, further solves the above formula (3), introduces the quadratic penalty factor α and the Lagrangian operator λ(t), and uses the following formula (4) to solve the nonlinear constraint problem and further obtain the optimal solution:

[0124]

[0125] In the formula, f(t) represents the historical power load data, and L() represents the expression of the introduced Lagrangian operator. In order to find the optimal solution to the original problem, the alternating direction method of multipliers (ADMM) algorithm is introduced to iteratively solve formula (4), and all modal components are converted to frequency domain calculation to obtain formula (5):

[0126]

[0127] In the formula, Remaining quantity The result after the Wiener filter is: represents the Fourier transform of f(t), Indicates u kThe Fourier transform of (t), ω represents the frequency, represents the Fourier transform of the Lagrangian operator λ(t), n is the number of iterations, Represents the center frequency of the modal component of the nth iteration. According to the same steps, the frequency domain analysis is performed to obtain the update formula of the center frequency of each modal component as shown in formula (6):

[0128]

[0129] In the formula, represents the center frequency of the modal component of the n+1th iteration, and ω is a parameter related to the center frequency. By setting the modal number k, repeating the above process, the VMD decomposition mode, i.e., k historical power data modal components, can be obtained.

[0130] Through VMD decomposition, we can obtain multiple modal components (IMF) with obvious regularity. In order to avoid the problem of spectrum aliasing in traditional decomposition, the number of decompositions is selected as 5 (only as an example) based on the center frequency observation method. The VMD decomposition results are as follows: Figure 2 As shown in the figure. IMF1 retains the fluctuation trend in the original sequence, so it has large volatility and no obvious regularity. IMF2-IMF5 can be seen to have strong periodicity, so the regularity is good. In addition, the frequency domain analysis is used to draw the spectrum of each modal component as shown in the figure. Figure 3 As shown, it can be seen that VMD decomposition does solve the problem of spectrum aliasing.

[0131] After each modal component of historical power data passes through the convolutional neural network module, i.e., the CNN layer, an output matrix can be calculated, which is the historical power load data feature. In the embodiment of the present invention, the 1D-CNN layer is selected when building the convolutional neural network model. The basic principle of 1D-CNN is as follows: Figure 4 As shown in the figure, a set of convolution kernels are used to perform convolution operations on the input sequence to generate a set of new feature sequences. Convolution kernels are usually small, learnable matrices that can detect different patterns and features in the input sequence. By applying convolution kernels at different positions in the input sequence, 1D-CNN can capture local patterns and structures in the sequence. After the convolution operation, 1D-CNN usually uses a pooling layer to reduce the number of features. A common pooling operation is maximum pooling, which extracts the maximum value from the local area of ​​each feature sequence, thereby retaining the most important feature information. Finally, the feature sequence is classified or regressed through a fully connected layer. 1D-CNN performs well in processing sequence data and can be used in fields such as text classification, speech recognition, and genomics. Their advantages include processing variable-length sequences, capturing local structures, and automatic feature extraction.

[0132] The output matrix of the convolutional neural network module is input to the long short-term memory network module (LSTM layer). By selecting key model parameters, future short-term power forecast data can be calculated. The basic structure of the LSTM layer is as follows: Figure 5 As shown in the figure, c t is the current cell state, h t is the hidden state, i t 、f t , O t They are input gate, forget gate, and output gate. t is the input of the cell, Z t is a temporary cell state, c t-1 is the previous cell state, h t-1 is the previous cell hidden state, h t+1 is the next cell hidden state, x t-1 is the input of the previous cell, x t+1 is the input of the next cell unit, σ and tanh are activation functions, and W f , W i , W O , W Z 、b f 、b i 、b O 、b g All gate training parameters participate in training. The LSTM network continuously learns and updates through the following formulas, where the forget gate structure saves the hidden state of the previous cell to the current cell, the input gate saves the current input information to the current cell, and the output gate saves the current cell state to the next cell. The output of the last cell is the future short-term power forecast data.

[0133] f t =σ(W f *[h t-1 ,x t ]+b f )

[0134] i t =σ(W i *[h t-1 ,x t ]+b i )

[0135] O t =σ(W O *[h t-1 ,x t ]+b O )

[0136] Z t =tanh(W Z *[ht-1 ,x t ]+b g )

[0137]

[0138]

[0139] In the embodiment of the present invention, when building the long short-term memory network module, the BiLSTM network is introduced to perform actual prediction work on power data. The advantage of the bidirectional network is that the introduction of the bidirectional structure allows the model to learn the intrinsic relationship of the data from the past and future dimensions of the time series, which can effectively improve the prediction ability and quantity utilization of the model. The structure of the bidirectional network BiLSTM is as follows: Figure 6 As shown in the figure, the combination process of each hidden layer can be represented by a combination of multiple formulas. In the figure, LSTM (Long Short-Term Memory) represents the operation process of LSTM network / LSTM layer.

[0140] In one implementation, the BiLSTM layer in the initial short-term power forecasting model is used to perform load forecasting on the historical power load data characteristics to obtain future short-term power forecasting data, including: using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain multiple load forecasting results; summing up the multiple load forecasting results to obtain future short-term power forecasting data. In this implementation, all predicted values ​​are summed up to restore the power data forecast output. For example, each power data forecast output sequence is recorded as y(t) m , if the first power modal component prediction output is recorded as y(t)1, the final prediction sequence y(t) is Where m is the number of power modal components.

[0141] From the above description, it can be seen that the flowchart of the VMD-CNN-BiLSTM model used in the short-term power forecasting model in the embodiment of the present invention is as follows: Figure 7As shown. After obtaining the characteristic data set about the power load (that is, the above-mentioned historical power load data set), the data is first normalized by the maximum and minimum standardization method; then the data is preprocessed, which includes converting the load data set into a supervised learning sequence, selecting a specific input duration and corresponding data features; decomposing it into a specific number of IMF modal components through VMD; each modal component is input into the 1D-CNN layer to extract data features; load forecasting is performed through the BiLSTM layer; finally, the forecast results are summed to evaluate the model results. In this embodiment, by combining the mining ability of CNN and the time series forecasting ability of BiLSTM, a network structure of a CNN-BiLSTM hybrid deep learning network is established. One-dimensional convolutional neural network (1D-CNN) is a neural network based on convolutional layers for processing one-dimensional sequence data, such as time series or text data. Unlike traditional fully connected neural networks, 1D-CNN uses convolution kernels to extract local features and compresses these features into less information through pooling layers. In the prediction training of traditional LSTM, the power load data is input and back-propagated according to the order of the time series. This method leads to a low utilization rate of the time series itself, and the model cannot fully learn the deeper internal relationship of the data. Therefore, the BiLSTM network is introduced to perform actual prediction work on power data. The advantage of the bidirectional network is that the introduction of a bidirectional structure allows the model to learn the internal relationship of the data from the past and future dimensions of the time series, which can effectively improve the prediction ability and quantity utilization of the model.

[0142] The parameter setting of the initial short-term power forecasting model is completed by manual parameter adjustment. The training set and the test set are divided according to 8:2. After the model training is completed, the test set data is used to perform single-point data prediction and obtain the prediction sequence. Since the maximum-minimum normalization is performed before the data is input, the prediction sequence should be denormalized, and the difference between the predicted value of the prediction sequence and the actual value of the original actual sequence should be compared and analyzed to evaluate the reliability of the model. It is stipulated that the prediction of the sample point whose prediction value is within 1% of the actual value is accurate, and the prediction accuracy is calculated. In an example, the deep learning model used in this scheme is constructed by tensor flow (an algorithm suitable for machine learning), the training rounds are set according to the actual experimental conditions, the relevant parameters are adjusted according to the actual work needs, and the loss function selects the root mean square error combined with the adaptive moment estimation (Adaptive Moment Estimation, Adam) optimizer and is optimized and adjusted in combination with the specific data set to obtain the final short-term power forecasting model.

[0143] The method for constructing the atmospheric pollutant emission prediction model in S103 above includes:

[0144] Obtain pollutant emission labeling data sets, which include historical power load data and pollutant emission monitoring data for the same period;

[0145] The PSO algorithm is used to optimize the initial hyperparameters of the SVR model to obtain the optimal hyperparameters;

[0146] The historical power load data and the pollutant emission monitoring data of the same period were used to iteratively train the SVR model with the optimal hyperparameters to obtain the atmospheric pollutant emission prediction model.

[0147] The pollutant emission monitoring data in the pollutant emission labeling data set can be the online monitoring data of the enterprise pollutants obtained as the target value, which plays the role of data labeling in the data set. The process of obtaining the pollutant emission labeling data set can include first collecting and collating the data to obtain the power-pollutant database, and then analyzing the database to obtain the pollutant emission labeling data set.

[0148] For example, in the process of data collection and collation, the power data used comes from the State Grid Information Center, which is collected from the user's smart meter, and the collection frequency is once every 15 minutes. The power data has been cleaned by the above data cleaning method, and its users cover 1,000 key industrial enterprises in a certain city. The pollutant data comes from the online monitoring system, covering a total of 127 key enterprises. The main monitoring pollutant indicators include sulfur dioxide, nitrogen oxides, smoke, etc. The data cleaning method is the same as the power data cleaning method. Through the precise matching of enterprise names, an electric power-environmental protection fusion database is established, which contains enterprise electricity consumption and enterprise pollutant information for a total of 50 enterprises. By collecting public information, daily-scale temperature data, working day data, and control day data are obtained. The above data granularity is not consistent, and the daily granularity is considered as the research granularity. The data set is divided into training set, test set, and validation set in a ratio of 7:2:1. In the process of analyzing the power-pollutant database, data is extracted from the database, and the industry and enterprise conduct power-pollutant Pearson correlation analysis, and the industry average power-pollution correlation coefficient r is calculated respectively. The value range of the Pearson correlation coefficient r: |r|≥0.70 indicates strong correlation, 0.30≤|r|≤0.70 indicates medium correlation, and |r|≤0.30 indicates weak correlation. The results show that the power data and pollutant data are not simple and intuitive linear relationships, and the power data and pollutant data cannot be directly modeled using a linear model. Seeking a nonlinear model that can adaptively fit power data and pollutant data is a possible breakthrough direction for pollutant data simulation technology. In an embodiment of the present invention, a machine learning algorithm can be used to construct a nonlinear correlation between power data and pollutant data, that is, a suitable model is preferably used to simulate the behavior of deriving pollutant data from power data, and then the accuracy of data derivation is gradually improved through model training, so that the model finally obtains accurate and reliable "power-to-pollution" capabilities, and directly characterizes pollutant emissions by changes in power data. Through these two steps, the model input and model output can be constructed, and the data can also be standardized to eliminate the dimensional influence between different features and improve the training effect of the model.

[0149] In the embodiment of the present invention, the atmospheric pollutant emission prediction model is an atmospheric pollutant emission data prediction module based on power data built by the PSO-SVR algorithm. Considering the existing data volume analysis, the number of data is relatively small, and the data samples of various industries are unevenly distributed. When selecting a model, it should have a certain ability to process small sample data; secondly, the model can meet the needs of sample differences and automatically optimize the model parameters to meet the pollutant data simulation of different users. Therefore, the PSO-SVR model is selected for analysis and modeling. SVR is formed on the basis of support vector machine classification (Support Vector Machines, SVM). Its essence is to find the optimal classification surface through training, so that the classification (prediction) error of all samples is minimized. The advantages of SVR are that it is suitable for regression problems of small number of sample data, the mathematical foundation is perfect, and it can also handle high-dimensional vector problems. PSO is a general optimization method with the advantages of simple method, fast convergence speed, and few parameter requirements. Therefore, it is often used in the optimization and parameter search process of many deep models. In one implementation, the PSO algorithm is used to optimize the initial hyperparameters of the SVR model to obtain the optimal hyperparameters, including: initializing the position and speed of each particle in the PSO algorithm, each particle represents a set of initial hyperparameters of the SVR model; calculating the fitness value of the position of each particle according to the fitness function, updating the speed and position of the particle, obtaining a new individual optimal position and a global optimal position of the particle group; iteratively updating the speed and position of the particle until the iteration termination condition is reached, and the particle with the optimal fitness value when the iteration termination condition is reached is used as the optimal hyperparameter of the SVR model. For example, the iteration termination condition includes obtaining the maximum number of iterations, or the optimal fitness does not change in a set number of consecutive iterations.

[0150] For example, when the particle swarm optimization algorithm (PSO) is initialized, a group of particles are randomly initialized, and each particle represents a set of parameters (C, γ) of the SVR model. The position of the particle represents the value of the parameter, and the velocity represents the direction and rate of change of the parameter. Set the size of the particle swarm to N and the maximum number of iterations to T. Set a random initial position and velocity for each particle. Evaluate the fitness of each particle, that is, calculate the root mean square error (RMSE) value of each particle according to the fitness function. Store the current position and fitness of each particle as its individual historical optimal solution pbest and the corresponding fitness f(pbest). Store the position and fitness of the particle with the best fitness among all particles as the global optimal solution gbest and the corresponding fitness f(gbest). Update the velocity and position of the particle. For each particle i, update its velocity V according to the following formula i and position x i :

[0151] Vi (t+1)=ω*V i (t)+c1*r1*(pbest-x i (t))+c2*r2*(gbest-x i (t))

[0152] x i (t+1)=x i (t)+V i (t+1)

[0153] Where ω is the inertia weight, c1 and c2 are learning factors, r1 and r2 are random numbers between [0, 1], and V i (t), V i (t+1) are the particle velocities at the tth iteration and the t+1th iteration, respectively, and x i (t), x i (t+1) are the particle positions at the tth iteration and the t+1th iteration, respectively. Check whether the termination condition is met. If the maximum number of iterations T is reached, or the value of the fitness function does not change significantly in several consecutive iterations, the iteration is stopped; otherwise, the initial parameter setting is returned. The parameters (C, γ) represented by the particle with the best fitness are selected as the final parameters of the SVR model to establish the final SVR model. Use all samples in the training set to train the final SVR model to obtain the final regression model, that is, the final atmospheric pollutant emission prediction model. In this example, the PSO model and the SVR model are run together. Specifically, the PSO algorithm is used as the (C, γ) data group input of the SVR model, and the SVR model inputs the power data for training, and then the SVR model with the best parameters is selected. The value of the fitness function can be calculated based on the power load data input to the SVR model, the pollutant prediction value output by the SVR model of the parameters corresponding to the particles in the PSO algorithm, and the pollutant monitoring value corresponding to the power load data.

[0154] In one example, in this embodiment, the model parameters are set according to relevant experience: the particle swarm size is set to 20, the self-weight is set to 1, the factors c1 and c2 are uniformly set to 2, and the maximum number of iterations is 30 times. Set the speed and initial position of each particle, and use the mean square error as the fitness function to calculate the fitness. According to the particle search situation, update the particle optimal solution and the global optimal solution, and obtain the current fitness value after iteration. Compare this value with the previous historical best value. If this value is better, replace the historical best value, otherwise perform the next calculation. Adjust the particle speed and position, and the algorithm goes through the above process several times, but when the iteration conditions are met, the calculation ends. Finally, the optimal parameter C and kernel function coefficient gamma under the above conditions can be returned.

[0155] The process of forecasting enterprise atmospheric pollutant emissions based on VMD-CNN-BiLSTM-PSO-SVR includes: a short-term power forecasting module based on VMD-CNN-BiLSTM deep learning model, and an atmospheric pollutant emission forecasting module based on power data built with PSO-SVR algorithm. The two modules are connected in series, and the short-term power forecasting data of the short-term power forecasting module is input into the atmospheric pollutant emission forecasting model to obtain the atmospheric pollutant emission forecasting results, that is, the simulation of future pollutant emissions. During the training process, verification and comparison results can be made based on the atmospheric pollutant emission forecasting results.

[0156] A specific embodiment is used to illustrate the following. Figure 8 As shown in the figure, characteristic data sets about power load and online monitoring data of pollutants are obtained, the power load data are normalized, preprocessed, and VMD decomposed, the pollution data are sorted and cleaned, and a power-environmental protection fusion database is established; then the VMD-CNN-BiLSTM short-term power forecasting model is trained, and the PS0-SVR pollution forecasting model is iteratively optimized; by inputting the company's historical power data, the short-term power forecasting model is run to obtain the short-term power load forecasting results, which drives the PSO-SVR pollution forecasting model, and finally the company's atmospheric pollutant emission forecasting results are obtained.

[0157] In the process of simulating future pollutants, the existing air quality simulation, prediction and early warning mostly use the method of simulating the pollutant values ​​of the current day and the next 5 days to forecast air concentration. In this embodiment, VMD-CNN-BiLSTM is used to predict the hourly load of the next 6 days and add them up to the daily electricity, and then the pollutant values ​​are fitted based on the PSO-SVR model. It is stipulated that the prediction accuracy of the sample points at the 6 sample points is within 10% of the actual value, and the prediction accuracy of the four industry data sets A, B, C, and D is calculated and evaluated. The sulfur dioxide emission prediction results are as follows: Fig. 9 As shown in Figure 2, the accuracy of the three data sets reached 83.33%, and the highest accuracy reached 100%. The prediction results of nitrogen oxide emissions are shown in Figure 2. Fig.10 As shown in Figure 2, the prediction accuracy can reach as low as 66.66% and as high as 83.33%. Fig.11 As shown, the prediction accuracy can be as low as 83.33% and as high as 100%. It can be seen that in the simulation process of three pollutant emissions, the overall prediction accuracy of each data set is above 66.66%, usually between 83.33% and 100%. The prediction results have a small deviation from the actual values, and the curve fitting is good. This model can be widely used in the prediction process of atmospheric pollutant emissions of enterprises. Furthermore, this technical method can predict the emission of related pollutants based on high-precision power data, and then locate and monitor key pollution sources, which has positive significance in the field of atmospheric pollutant emission control.

[0158] The current prediction methods are mainly based on the time series prediction of pollutants, that is, based on the activity level of pollutants in the past, relying on the efficient fitting of major models or algorithms to achieve the prediction of future activity levels. Specifically, it is usually to combine smart grids with big data, and to construct a pollution emission prediction method by analyzing the correlation between the electricity consumption of industrial enterprises and the total industrial output and the direct emissions of major pollutants. Deep learning models such as LSTM recurrent neural networks, ResNet and convolutional long short-term memory networks are used to improve the prediction accuracy of PM2.5 and other pollutant concentrations.

[0159] The existing technical solutions are:

[0160] (1) Prediction method based on statistical model: This method uses historical pollutant emission data and meteorological data to establish a prediction model through statistical analysis. However, this method is often limited by data quality and quantity, and the prediction accuracy needs to be improved.

[0161] (2) Prediction methods based on machine learning: In recent years, machine learning technology has been widely used in the prediction of atmospheric pollutants. By collecting and analyzing data such as enterprise electricity consumption and environmental parameters, a prediction model is established using machine learning algorithms (such as support vector machines, neural networks, random forests, etc.) to predict pollutant emissions. This method has high prediction accuracy, but requires a large amount of data and computing resources.

[0162] (3) Enterprise pollution prediction method based on power big data: This method collects the historical electricity consumption data and historical pollution data of enterprises, performs data preprocessing and feature extraction, and then builds machine learning models such as logistic regression models for prediction. This method combines power big data and machine learning technology to predict the pollution discharge of potential polluting enterprises and provide strong support for environmental protection.

[0163] Through the above description of the prior art, it can be seen that the prior art is a regression prediction model represented by statistics and machine learning for pollutant prediction. Such models need to collect a large amount of independent variable and dependent variable data, and establish models based on the objective relationship between various influencing factors and pollutant emissions. The defect of the above main technology is that in the process of establishing the enterprise pollutant prediction model, a large amount of relevant variable data needs to be collected to participate in the modeling, and the pollutant data is often the total emission data of the enterprise provided by the enterprise's online monitoring data, and the pollutant prediction model has poor accuracy and real-time performance. There are two specific problems. The first problem is that the computing resources are consumed: the model is complex → the amount of data processing is large → the computing resources are consumed: the prediction method based on machine learning requires a large amount of data and computing resources, especially for complex models and data sets, the training time and computing cost are high. The second problem is that the model has limited generalization ability: the enterprise conditions change → the model has poor adaptability → the prediction accuracy is reduced: although the enterprise pollution prediction method based on power big data combines power big data and machine learning technology, the generalization ability of the model is still subject to certain limitations. If the production process, equipment status or environmental conditions of the enterprise change, the prediction accuracy of the model may be affected.

[0164] In order to solve the above-mentioned pain points of the prior art and monitor the pollution emissions of enterprises more comprehensively and objectively, the present invention selects short-term prediction of power data and combines machine learning to build a pollutant regression model to achieve better prediction results for pollutants in the future. Then, the PSO-SVR algorithm is used to build an atmospheric pollutant emission data prediction model, the purpose of which is to achieve automatic adjustment of parameters to adapt to the ability to process data from enterprises in different industries, thereby driving pollutant emission prediction with power load prediction data, reducing the cost of atmospheric pollutant prediction for enterprises, improving the accuracy and real-time nature of prediction, and providing timely pollution prediction and decision-making support for enterprises and government departments, realizing differentiated management of different pollution sources, optimizing resource allocation, and improving management and control efficiency.

[0165] Specifically, in the VMD-CNN-BiLSTM power data short-term prediction model of the embodiment of the present invention, multiple modal components (IMFs) with obvious rules can be obtained through VMD decomposition, which effectively solves the problem of spectrum aliasing caused by high volatility and nonlinearity of power load data, thereby improving the prediction accuracy of the model. In the VMD-CNN-BiLSTM power data short-term prediction model, the BiLSTM network is used to perform actual prediction work on power data, allowing the model to learn the intrinsic relationship of data from the past and future dimensions of the time series in both directions, effectively improving the prediction ability and data utilization of the model, while ensuring the prediction accuracy of the model, the generalization ability of the model can be improved, and the prediction accuracy of the model can be guaranteed. The PSO-SVR algorithm is adopted to build an atmospheric pollutant emission data prediction model, and the advantage of SVR is suitable for regression problems of small number of sample data, which saves computing resources and computing costs. The PSO-SVR algorithm is adopted to build an atmospheric pollutant emission data prediction model, and the advantages of PSO method are simple, fast convergence speed, and few parameter requirements, which saves computing resources and computing costs; its powerful optimization parameter search function meets the sample difference needs of model construction of different enterprises, and ensures the prediction accuracy of the model. The enterprise atmospheric pollutant emission prediction method based on VMD-CNN-BiLSTM-PSO-SVR uses power big data to predict short-term power load, and uses power load prediction data to drive pollutant emission prediction. The accuracy of the two-level prediction model is guaranteed, and the advantage of power big data is transferred to future atmospheric pollutant prediction data, which improves the accuracy and real-time performance of the prediction, provides timely pollution prediction and decision-making support for enterprises and government departments, realizes differentiated management of pollution sources, optimizes resource allocation, and improves management and control efficiency.

[0166] In the embodiment of the present invention, the short-term prediction model of power data uses VMD decomposition to obtain multiple (for example, 5) modal components (IMFs) with obvious rules. For time series prediction based on single user variables, it is an effective means to solve the problem of spectrum aliasing. The prediction model of atmospheric pollutant emission data uses the PSO-SVR algorithm, which can save computing resources and computing costs while meeting the sample difference needs of model construction of different enterprises. The enterprise atmospheric pollutant emission prediction method based on VMD-CNN-BiLSTM-PSO-SVR predicts short-term power load with power big data, and drives pollutant emission prediction with power load prediction data. The accuracy of the two-level prediction model is guaranteed, realizing the transmission of the advantages of power big data to future atmospheric pollutant prediction data, improving the accuracy and real-time performance of the prediction, and providing timely pollution prediction and decision-making support for enterprises and government departments, realizing differentiated management of pollution sources, optimizing resource allocation, and improving management and control efficiency.

[0167] The embodiment of the present invention relates to the electric power field and the field of air pollutant prevention and control, and is a method for predicting enterprise air pollutant emission data based on electric power big data. The main problem solved is how to effectively use the advantages of electric power big data such as strong real-time, high accuracy, high resolution, and wide collection range to achieve real-time and accurate prediction of enterprise air pollutant emissions. Specifically, use electric power data to drive the prediction of future air pollutant emission data, and realize data advantage transmission through model construction technologies such as data analysis and machine learning. Timely discover and prevent potential pollution incidents, and provide a scientific basis for environmental protection and pollution control.

[0168] Embodiment 2:

[0169] Based on the same inventive concept, the present invention also provides an enterprise atmospheric pollutant emission prediction system based on power load data, the structural diagram of which is shown in FIG. Fig.12 As shown, including:

[0170] An acquisition module is used to obtain the current power load data of the enterprise;

[0171] A load forecasting module, used for inputting the current power load data into a short-term power forecasting model to perform load forecasting and obtain current short-term power forecasting data;

[0172] A pollutant emission prediction module, used to input the current short-term power prediction data into the atmospheric pollutant emission prediction model for prediction, and obtain the atmospheric pollutant emission prediction result of the enterprise;

[0173] The short-term power forecasting model is composed of a variational mode decomposition (VMD) algorithm, a convolutional neural network (CNN) layer, and a bidirectional long short-term memory (BiLSTM) network layer; the atmospheric pollutant emission forecasting model is constructed based on a particle swarm optimization (PSO) algorithm and a support vector machine regression (SVR) model.

[0174] In a possible implementation, a short-term power forecasting model building module is further included, which is used to:

[0175] Obtain historical power load data sets;

[0176] Using a time series construction method, historical power load data of a specific duration and future short-term power data corresponding to the historical power load data are selected from the historical power load data set;

[0177] The VMD algorithm in the initial short-term power forecasting model is used to perform signal decomposition on the historical power load data to obtain multiple modal components of the historical power data;

[0178] Using the one-dimensional CNN layer in the initial short-term power forecasting model to extract features of the multiple historical power data modal components to obtain historical power load data features;

[0179] Using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain future short-term power forecasting data;

[0180] Based on the future short-term power data and the future short-term prediction data, the initial short-term power prediction model is iteratively trained to obtain a final short-term power prediction model.

[0181] In a possible implementation, the short-term power forecasting model building module is specifically used to:

[0182] Normalizing the historical power load data set using a maximum-minimum normalization algorithm to obtain a normalized historical power load data set;

[0183] Using a time series construction method to convert the normalized historical power load data set into a supervised learning sequence;

[0184] In the supervised learning sequence, historical power load data of a specific duration and future short-term power data corresponding to the historical power load data are selected.

[0185] In a possible implementation, the short-term power forecasting model building module is specifically used to:

[0186] Using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain multiple load forecasting results;

[0187] The multiple load forecast results are summed to obtain future short-term power forecast data.

[0188] In a possible implementation, the short-term power forecasting model building module is further used to:

[0189] Obtain historical power load forecast sets;

[0190] Inputting the historical power load forecast set into the short-term power forecast model to perform load forecasting, and obtaining a power forecast data sequence of the historical power load forecast set;

[0191] Performing a denormalization process on the power prediction data sequence to obtain a denormalized power prediction data sequence;

[0192] The short-term power forecasting model is evaluated based on the real power data sequence in the historical power load forecasting set and the denormalized power forecasting data sequence.

[0193] In a possible implementation, the short-term power forecasting model building module is further used to:

[0194] Analyzing the characteristic correlation of the characteristic parameters in the normalized historical power load data set using the Pearson coefficient analysis method to obtain the correlation coefficient of each characteristic parameter;

[0195] Based on each characteristic parameter, if the correlation coefficient of the characteristic coefficient exceeds the correlation coefficient threshold, the historical power load data of the characteristic parameter is saved in the historical power load data set, otherwise the historical power load data of the characteristic parameter is deleted from the historical power load data set.

[0196] In a possible implementation, an atmospheric pollutant emission prediction model building module is further included, which is used to:

[0197] Obtaining a pollutant emission labeling data set, wherein the pollutant emission labeling data set includes historical power load data and pollutant emission monitoring data of the same period;

[0198] The PSO algorithm is used to optimize the initial hyperparameters of the SVR model to obtain the optimal hyperparameters;

[0199] The historical power load data and the pollutant emission monitoring data of the same period are used to iteratively train the SVR model with the optimal hyperparameters to obtain an atmospheric pollutant emission prediction model.

[0200] In a possible implementation, the atmospheric pollutant emission prediction model construction module is specifically used to:

[0201] Initialize the position and velocity of each particle in the PSO algorithm. Each particle represents a set of initial hyperparameters of the SVR model.

[0202] The fitness value of each particle's position is calculated according to the fitness function, and the particle's speed and position are updated to obtain a new individual optimal position and a global optimal position of the particle group. The particle's speed and position are updated repeatedly until the iteration termination condition is reached. The particle with the best fitness value when the iteration termination condition is reached is used as the optimal hyperparameter of the SVR model.

[0203] In a possible implementation manner, the iteration termination condition includes obtaining a maximum number of iterations, or the optimal fitness does not change in a set number of consecutive iterations.

[0204] Embodiment 3:

[0205] like Fig.13As shown, the present invention also provides an electronic device, which may be a computer device, a single-chip device, an intelligent mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected via a bus; the memory may be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory may also be used to store data, which may be called and / or modified when the instructions are executed.

[0206] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in a storage medium to implement the corresponding method flow or corresponding functions, so as to implement the steps of a method for predicting enterprise atmospheric pollutant emissions based on power load data in the above-mentioned embodiment.

[0207] Embodiment 4:

[0208] Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It can be understood that the storage medium here can include both built-in storage media in electronic devices and, of course, extended storage media supported by electronic devices. The storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by a processor are also stored in the storage space, and these instructions can be one or more execution programs (including program codes). It should be noted that the storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage. The processor loads and executes one or more instructions stored in the storage medium, which can implement the steps of a method for predicting enterprise atmospheric pollutant emissions based on power load data in the above embodiment.

[0209] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0210] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0211] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0212] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0213] 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 its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims to be approved.

Claims

1. A method for predicting enterprise atmospheric pollutant emissions based on power load data, characterized in that: include: Obtain the current power load data of the enterprise; Inputting the current power load data into a short-term power forecasting model to perform load forecasting to obtain current short-term power forecasting data; Inputting the current short-term power forecast data into an air pollutant emission forecast model for forecasting to obtain an air pollutant emission forecast result of the enterprise; The short-term power forecasting model is composed of a variational mode decomposition (VMD) algorithm, a convolutional neural network (CNN) layer, and a bidirectional long short-term memory (BiLSTM) network layer; the atmospheric pollutant emission forecasting model is constructed based on a particle swarm optimization (PSO) algorithm and a support vector machine regression (SVR) model.

2. The method according to claim 1, characterized in that The construction process of the short-term power forecasting model includes: Obtain historical power load data sets; Using a time series construction method, historical power load data of a specific duration and future short-term power data corresponding to the historical power load data are selected from the historical power load data set; The VMD algorithm in the initial short-term power forecasting model is used to perform signal decomposition on the historical power load data to obtain multiple modal components of the historical power data; Using the one-dimensional CNN layer in the initial short-term power forecasting model to extract features of the multiple historical power data modal components to obtain historical power load data features; Using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain future short-term power forecasting data; Based on the future short-term power data and the future short-term prediction data, the initial short-term power prediction model is iteratively trained to obtain a final short-term power prediction model.

3. The method according to claim 2, characterized in that The time series construction method is used to select historical power load data of a specific time length and future short-term power data corresponding to the historical power load data from the historical power load data training set, including: Normalizing the historical power load data set using a maximum-minimum normalization algorithm to obtain a normalized historical power load data set; Using a time series construction method to convert the normalized historical power load data set into a supervised learning sequence; In the supervised learning sequence, historical power load data of a specific duration and future short-term power data corresponding to the historical power load data are selected.

4. The method according to claim 2 or 3, characterized in that The method of using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain future short-term power forecasting data includes: Using the BiLSTM layer in the initial short-term power forecasting model to perform load forecasting on the historical power load data characteristics to obtain multiple load forecasting results; The multiple load forecast results are summed to obtain future short-term power forecast data.

5. The method according to claim 3, characterized in that After the initial short-term power prediction model is iteratively trained based on the future short-term power data and the future short-term prediction data to obtain a final short-term power prediction model, the method further includes: Obtain historical power load forecast sets; Inputting the historical power load forecast set into the short-term power forecast model to perform load forecasting, and obtaining a power forecast data sequence of the historical power load forecast set; Performing a denormalization process on the power prediction data sequence to obtain a denormalized power prediction data sequence; The short-term power forecasting model is evaluated based on the real power data sequence in the historical power load forecasting set and the denormalized power forecasting data sequence.

6. The method according to claim 3, characterized in that After the historical power load data set is normalized by using the maximum and minimum normalization algorithm to obtain the normalized historical power load data set, the method further includes: Analyzing the characteristic correlation of the characteristic parameters in the normalized historical power load data set using the Pearson coefficient analysis method to obtain the correlation coefficient of each characteristic parameter; Based on each characteristic parameter, if the correlation coefficient of the characteristic coefficient exceeds the correlation coefficient threshold, the historical power load data of the characteristic parameter is saved in the historical power load data set, otherwise the historical power load data of the characteristic parameter is deleted from the historical power load data set.

7. The method according to claim 1, characterized in that The method for constructing the atmospheric pollutant emission prediction model comprises: Obtaining a pollutant emission labeling data set, wherein the pollutant emission labeling data set includes historical power load data and pollutant emission monitoring data of the same period; The PSO algorithm is used to optimize the initial hyperparameters of the SVR model to obtain the optimal hyperparameters; The historical power load data and the pollutant emission monitoring data of the same period are used to iteratively train the SVR model with the optimal hyperparameters to obtain an atmospheric pollutant emission prediction model.

8. The method according to claim 7, characterized in that The PSO algorithm is used to optimize the initial hyperparameters of the SVR model to obtain the optimal hyperparameters, including: Initialize the position and velocity of each particle in the PSO algorithm. Each particle represents a set of initial hyperparameters of the SVR model. The fitness value of each particle's position is calculated according to the fitness function, and the particle's speed and position are updated to obtain a new individual optimal position and a global optimal position of the particle group. The particle's speed and position are updated repeatedly until the iteration termination condition is reached. The particle with the best fitness value when the iteration termination condition is reached is used as the optimal hyperparameter of the SVR model.

9. The method according to claim 8, characterized in that The iteration termination condition includes obtaining a maximum number of iterations, or the optimal fitness does not change in a set number of consecutive iterations.

10. A system for predicting enterprise air pollutant emissions based on power load data, characterized in that: include: An acquisition module is used to obtain the current power load data of the enterprise; A load forecasting module, used for inputting the current power load data into a short-term power forecasting model to perform load forecasting and obtain current short-term power forecasting data; A pollutant emission prediction module, used for inputting the current short-term power prediction data into the atmospheric pollutant emission prediction model for prediction, and obtaining the atmospheric pollutant emission prediction result of the enterprise; The short-term power forecasting model is composed of a variational mode decomposition (VMD) algorithm, a convolutional neural network (CNN) layer, and a bidirectional long short-term memory (BiLSTM) network layer; the atmospheric pollutant emission forecasting model is constructed based on a particle swarm optimization (PSO) algorithm and a support vector machine regression (SVR) model.

Images