A method, system, electronic device and storage medium for predicting comprehensive power consumption of power system based on deep learning

Abstract

The present application provides a method, system, electronic device, and storage medium for predicting comprehensive plant power consumption of an electric power system based on deep learning, wherein the prediction method comprises: collecting feature data at each moment in a prediction area and constructing a feature data set; encoding the feature data set using an improved autoencoder to obtain encoded time series features, and reconstructing the encoded time series features using a decoder to obtain reconstructed time series features; judging whether the reconstructed time series features are consistent with the feature data set, and if not, adjusting the improved autoencoder and decoder through a loss function until the adjusted reconstructed time series features are consistent with the feature data set; processing the adjusted reconstructed time series features through a long short-term memory network to obtain prediction results of the comprehensive plant power consumption rate and power load of the electric power system; having the beneficial effect of being able to efficiently and accurately predict the comprehensive plant power consumption rate and power load; and being applicable to the field of comprehensive plant power consumption prediction of electric power systems.

Description

Technical Field

[0001] The present application relates to the field of comprehensive plant power consumption prediction of power systems, and specifically to a method, system, electronic equipment, and storage medium for comprehensive plant power consumption prediction of power systems based on deep learning. Background Art

[0002] The main idea of ​​comprehensive power load forecasting in power systems is to effectively analyze and forecast based on the changing patterns of regional power load. Traditional power load forecasting mainly uses methods such as unit consumption method, elastic coefficient method, and load density method. With the development of social economy, my country's energy has undergone transformation, and the production structure is constantly being adjusted, making the characteristics of loads more complex. Load forecasting has begun to generally use methods such as sequence recursion and linear regression, as well as artificial intelligence technologies such as expert systems and neural networks to identify the laws of loads.

[0003] In the new power system, a large number of new energy sources are connected to the grid, both distributed and centralized, which has a great impact on the prediction of regional network supply (centralized distribution) load, affecting the accuracy of regional load prediction, and bringing great challenges to regional load prediction. At the same time, a large number of distributed photovoltaics are connected to the medium and low voltage distribution network, the local balance of medium and low voltage loads and the transmission of distributed photovoltaic power have led to a significant underestimation of regional loads during the lunch peak, making regional load prediction more difficult.

[0004] With the advancement of new energy power generation prediction technology, regional new energy has gradually realized short-term, efficient and accurate prediction of new energy power generation based on comprehensive information such as meteorology, geographical location and different types of units, providing a reliable data basis for load prediction of new power systems. Compared with the load on the power consumption side, the change law of power generation load on the power generation side has scientific and reasonable analysis characteristics and strong interpretability. Therefore, it has become a new idea to predict the load of the whole region from the power generation side. However, in the process of exploring and applying this idea, due to the fundamental change in the power generation structure of the new power system, the power consumption of new energy units is completely different from that of traditional thermal power, hydropower and other units. Therefore, when the power consumption load is restored and converted through the power generation load, the change in the power consumption rate makes it difficult to restore the power generation and consumption loads. Due to the randomness and intermittency of the output of new energy, the output distribution of various types of units fluctuates greatly, which increases the difficulty of predicting the power consumption rate. Summary of the invention

[0005] In order to solve one of the above-mentioned technical defects, the present application provides a method, system, electronic device and storage medium for predicting comprehensive power consumption of power system based on deep learning.

[0006] According to a first aspect of the present application, a method for predicting comprehensive power consumption of a power system based on deep learning is provided, comprising:

[0007] Collect characteristic data at each moment in the forecast area, and construct a characteristic data set based on the characteristic data at each moment; wherein the characteristic data includes the planned output of each unit and the planned load of the entire area;

[0008] An improved automatic encoder is used to encode the feature data set to obtain encoded time series features, and then a decoder is used to reconstruct the encoded time series features to obtain reconstructed time series features;

[0009] Determine whether the reconstructed time series features are consistent with the feature data set. If not, adjust the improved automatic encoder and decoder through the loss function until the adjusted reconstructed time series features are consistent with the feature data set.

[0010] The adjusted reconstructed time series features are processed through long short-term memory networks to obtain the prediction results of the comprehensive plant power consumption rate and power load of the power system.

[0011] Preferably, constructing a feature data set according to feature data at each moment specifically includes:

[0012] The corresponding time feature vector is constructed based on the characteristic data collected at each time in the prediction area; the feature vector x at time t t for:

[0013] x t =[G t ,S t ,W t ,F t ,P t ,D t ,L t ];

[0014] In the formula, G t represents the planned output of the thermal power unit at time t, S t represents the planned output of the hydropower unit at time t, W t represents the planned output of the wind turbine at time t, F t represents the planned output of the gas unit at time t, P t represents the planned output of the centralized photovoltaic power station at time t, D t represents the planned output of distributed photovoltaic power generation at time t, L t represents the planned load of the whole area at time t;

[0015] Construct the time series variation feature X at time t t For: X t =[x t-1 x t x t+1 ] T ; In the formula, x t-1 represents the eigenvector at time t-1, xt+1 Represents the feature vector at time t+1;

[0016] Construct feature data set: X = [X1, ..., X t ,……,X N ], where X N It represents the time series variation feature at N moments, t=1, 2, ..., N, where N represents the number of acquisitions, i.e. the number of moments.

[0017] Preferably, the improved autoencoder comprises three layers;

[0018] The improved automatic encoder is used to encode the feature data set to obtain the encoded time series features, which specifically includes:

[0019] The feature data set is input into the first layer of the improved autoencoder, and the interdependence features between the time series change feature quantities at each moment in the feature data set are extracted through the self-attention mechanism; the gated recurrent network is used to extract the overall time series features from the extracted interdependence features; the extraction result of the overall time series features is used as the output result of the first layer of the improved autoencoder;

[0020] Inputting the output result of the first layer of the improved autoencoder into the second layer of the improved autoencoder;

[0021] Inputting the output result of the second layer of the improved autoencoder into the third layer of the improved autoencoder;

[0022] The output result of the third layer of the improved autoencoder is used as the extracted encoded temporal features.

[0023] Preferably, the extraction formula of the self-attention mechanism is:

[0024]

[0025] In the formula, Z i represents the interdependency feature, Q represents the query matrix, Q = XW Q , K represents the key matrix, K = XW K , V represents the value matrix, V = XW V , d k represents the dimension of the query matrix Q or key matrix K, W Q , W K , W V Represent the weight coefficients of the query matrix, key matrix, and value matrix respectively.

[0026] Preferably, the extracting of overall temporal features from the extracted interdependence features using a gated recurrent network specifically includes:

[0027] Reset gate calculation, the formula for reset gate calculation is: R t =σ(Z i W xr +H t-1 W hr +b r );

[0028] In the formula, R t represents the output of the reset gate at time t, σ(·) represents the activation function, and Z i Represents the interdependence characteristics, H t-1 represents the hidden state at time t-1, W xr , W hr , b r Represents the learnable weight parameters of the reset gate;

[0029] Update gate calculation, the formula for update gate calculation is: Z t =σ(Z i W xz +H t-1 W hz +b z );

[0030] In the formula, Z t represents the output of the update gate at time t, W xz , W hz , b z Represents the learnable weight parameters of the update gate;

[0031] Under the action of the reset gate, the candidate hidden state is calculated. The calculation formula of the candidate hidden state is: H t Y =tanh(Z i W xh +(R t ⊙H t-1 )W hh +b h );

[0032] In the formula, H t Y represents the candidate hidden state at time t, W xh , W hh , b h represents the learnable weight parameter of the candidate hidden state, ⊙ represents the Hadamard product;

[0033] Hidden state calculation, the calculation formula of the hidden state is: H t =Z t ⊙H t-1 +(1-Z t )⊙H t Y ;

[0034] In the formula, H t represents the hidden state at time t;

[0035] The hidden state matrix is ​​constructed through the hidden state at each moment, and the hidden state matrix is ​​used as the extraction result of the overall time series feature.

[0036] Preferably, the adjusted reconstructed time series features are processed by a long short-term memory network to obtain the prediction results of the comprehensive plant power consumption rate and power load of the power system, specifically including:

[0037] Forget gate calculation, the calculation formula of the forget gate is: f t =sigmoid(W if A td +b if +W hf h t-1 +b hf );

[0038] In the formula, f t Represents the output of the forget gate at time t, A td represents the reconstructed timing characteristics after adjustment at time t, W if represents the weight matrix between the reconstructed time series features and the forget gate after adjustment at time t, b if represents the bias matrix between the reconstructed time series features adjusted at time t and the forget gate, W hf represents the weight matrix between the output gate and the forget gate at time t-1, b hf represents the bias matrix between the output gate and the forget gate at time t-1, h t-1 Represents the output result of the output gate at time t-1;

[0039] Input gate calculation, the calculation formula of the input gate is: c t =f t ⊙c t-1 +i t ⊙g t ;

[0040] In the formula, c t represents the output result of the input gate at time t, c t-1 represents the output result of the input gate at time t-1, i t represents the self-learning output result of the output layer feature of the input gate at time t, i t =sigmoid(W ii A td +b ii +W hi h t-1 +b hi ), W iirepresents the self-learning weight matrix of the reconstructed time series features after adjustment at time t, b ii W represents the bias matrix of the reconstructed time series features after adjustment at time t; hi represents the self-learning weight matrix of the output gate at time t-1, b hi represents the bias matrix of the output gate at time t-1, g t Indicates the output result of the input gate at time t based on the learning of the assigned weight matrix features, g t =tanh(W ig A td +b ig +W hg h t-1 +b hg ), W ig represents the weight matrix between the reconstructed timing features and the input gate after adjustment at time t, b ig represents the bias matrix between the reconstructed timing features and the input gate after adjustment at time t, W hg represents the weight matrix between the output gate and the input gate at time t-1, b hg Represents the bias matrix between the output gate and the input gate at time t-1;

[0041] Output gate calculation, the calculation formula of the output gate is: h t =o t ⊙tanh(c t );

[0042] In the formula, h t represents the output result of the output gate at time t, o t =sigmoid(W io A td +b io +W ho h t-1 +b ho ), W io represents the weight matrix between the reconstructed timing features and the output gate after adjustment at time t, b io represents the bias matrix between the reconstructed timing features and the output gate after adjustment at time t, W ho represents the weight matrix between the output gate at time t-1 and the output gate at time t, b ho Represents the bias matrix between the output gate at time t-1 and the output gate at time t;

[0043] The comprehensive plant power consumption rate η at time t is calculated by the first fully connected neural network t , η t =W o h t +b o Where W orepresents the weight matrix of the first fully connected neural network, b o represents the bias matrix of the first fully connected neural network;

[0044] The power load forecast result L after the power generation side is restored at time t is calculated by the comprehensive power consumption rate at time t t ′,L t ′=G t +S t +W t +F t +P t +D t (1-η t ).

[0045] Preferably, the second layer and the third layer of the improved autoencoder both use a fully connected neural network for feature extraction;

[0046] The calculation formula for feature extraction of the second layer of the improved autoencoder is: tq =W r1 H t +b r1 Where H tq represents the output of the second layer of the improved autoencoder at time t, W r1 represents the weight coefficient of the second layer of the improved autoencoder, H t represents the time series feature at time t in the overall time series feature, that is, the hidden state at time t, b r1 represents the bias coefficient of the second layer of the improved autoencoder;

[0047] The calculation formula for feature extraction of the third layer of the improved autoencoder is: tp =W r2 H tq +b r2 Where H tp represents the output of the third layer of the improved autoencoder at time t, W r2 represents the weight coefficient of the third layer of the improved autoencoder, b r2 Represents the bias coefficients of the third layer of the improved autoencoder.

[0048] According to a second aspect of the present application, a deep learning-based integrated power system plant power consumption prediction system is provided, comprising a module for implementing a deep learning-based integrated power system plant power consumption prediction method as described in any of the above items.

[0049] According to a third aspect of the present application, an electronic device is provided, including:

[0050] Memory;

[0051] Processor; and

[0052] Computer programs;

[0053] Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the deep learning-based comprehensive plant power consumption prediction method for the power system as described in any one of the above.

[0054] According to a fourth aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored; the computer program is executed by a processor to implement a method for predicting comprehensive plant power consumption of a power system based on deep learning as described in any one of the above items.

[0055] The present application provides a reliable and scientific method for predicting the comprehensive power consumption of power plants in power systems, which can realize the efficient and accurate prediction of the comprehensive power consumption rate and power load of power plants. Since the output and load of each unit are uncertain and change at all times, based on the planned output of each unit on the power generation side and the time series change of the planned output of each unit, the improved automatic encoder and decoder are used to process the feature data set to obtain the reconstructed time series features, and then the comprehensive power consumption rate of power plants is calculated through the long short-term memory network, and then the power generation side load restoration method is used to convert the power generation side to the power consumption side through the comprehensive power consumption rate, so as to realize the calculation of power load.

[0056] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the contents indicated in the written description and the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0058] Figure 1 A flowchart of a method for predicting power consumption of a power system based on deep learning provided in an embodiment of the present application;

[0059] Figure 2 for Figure 1 A schematic diagram of the encoding process of the improved automatic encoder and the decoding process of the decoder in the embodiment;

[0060] Figure 3 for Figure 1 A schematic diagram of the first layer of the improved automatic encoder in the embodiment;

[0061] Figure 4 for Figure 1 A schematic diagram of a process of processing the adjusted reconstructed time series features by a long short-term memory network in the embodiment;

[0062] Figure 5 A functional structure diagram of a power system comprehensive plant power consumption prediction system based on deep learning provided in an embodiment of the present application;

[0063] Figure 6 for Figure 5 A schematic diagram of the functional structure of a feature data set construction module in the embodiment;

[0064] Figure 7 for Figure 5 A schematic diagram of the functional structure of the encoding module in the embodiment;

[0065] Figure 8 for Figure 5 A functional structure diagram of the reconstruction module in the embodiment;

[0066] Fig. 9 for Figure 5 A schematic diagram of the functional structure of the prediction module in the embodiment;

[0067] In the figure:

[0068] 10 is an acquisition module, 20 is a feature data set construction module, 30 is an encoding module, 40 is a reconstruction module, 50 is a judgment module, 60 is an adjustment module, 70 is a prediction module, 201 is a moment feature vector construction unit, 202 is a time series change feature quantity construction unit, 203 is a feature data set construction unit, 301 is a self-attention mechanism extraction unit, 302 is a gated recurrent network extraction unit, 303 is a first processing unit, 304 is a second processing unit, 305 is a third processing unit, 306 is an encoding Time series feature extraction unit, 401 is the first reconstruction unit, 402 is the second reconstruction unit, 403 is the third reconstruction unit, 701 is the forget gate calculation unit, 702 is the input gate calculation unit, 703 is the output gate calculation unit, 704 is the comprehensive plant power consumption rate calculation unit, 705 is the power load calculation unit, 3021 is the reset gate calculation unit, 3022 is the update gate calculation unit, 3023 is the candidate hidden state calculation unit, 3024 is the hidden state calculation unit, and 3025 is the hidden state matrix construction unit. DETAILED DESCRIPTION

[0069] In order to make the technical solutions and advantages in the embodiments of the present application more clearly understood, the exemplary embodiments of the present application are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0070] Figure 1 A flowchart of a method for predicting power consumption of a power system based on deep learning is provided in an embodiment of the present application, such as Figure 1 As shown, in order to solve the above problems, an embodiment of the present application provides a method for predicting comprehensive power consumption of a power system based on deep learning, including:

[0071] Collect characteristic data at each moment in the forecast area, and construct a characteristic data set based on the characteristic data at each moment; wherein the characteristic data includes the planned output of each unit and the planned load of the entire area;

[0072] An improved automatic encoder is used to encode the feature data set to obtain encoded time series features, and then a decoder is used to reconstruct the encoded time series features to obtain reconstructed time series features;

[0073] Determine whether the reconstructed time series features are consistent with the feature data set. If not, adjust the improved automatic encoder and decoder through the loss function until the adjusted reconstructed time series features are consistent with the feature data set.

[0074] The adjusted reconstructed time series features are processed through long short-term memory networks to obtain the prediction results of the comprehensive plant power consumption rate and power load of the power system.

[0075] The present application provides a reliable and scientific method for predicting the comprehensive power consumption of power plants in power systems, which can realize the efficient and accurate prediction of the comprehensive power consumption rate and power load of power plants. Since the output and load of each unit are uncertain and change at all times, based on the planned output of each unit on the power generation side and the time series change of the planned output of each unit, the improved automatic encoder and decoder are used to process the feature data set to obtain the reconstructed time series features, and then the comprehensive power consumption rate of power plants is calculated through the long short-term memory network, and then the power generation side load restoration method is used to convert the power generation side to the power consumption side through the comprehensive power consumption rate, so as to realize the calculation of power load.

[0076] Furthermore, constructing a feature data set according to the feature data at each moment specifically includes:

[0077] The corresponding time feature vector is constructed based on the characteristic data collected at each time in the prediction area; the feature vector x at time t t for:

[0078] x t =[G t ,S t ,W t ,F t ,P t ,D t ,L t];

[0079] In the formula, G t represents the planned output of the thermal power unit at time t, S t represents the planned output of the hydropower unit at time t, W t represents the planned output of the wind turbine at time t, F t represents the planned output of the gas unit at time t, P t represents the planned output of the centralized photovoltaic power station at time t, D t represents the planned output of distributed photovoltaic power generation at time t, L t represents the planned load of the whole area at time t;

[0080] Construct the time series variation feature X at time t t For: X t =[x t-1 x t x t+1 ] T ; In the formula, x t-1 represents the eigenvector at time t-1, x t+1 Represents the feature vector at time t+1;

[0081] Construct feature data set: X = [X1, ..., X t ,……,X N ]; where X N It represents the time series variation feature at N moments, t=1, 2, ..., N, where N represents the number of acquisitions, i.e. the number of moments.

[0082] In this application, a characteristic data set is constructed according to the planned output of each unit, and the moment characteristic data of all units are collected from the power generation side to facilitate subsequent calculations using the power generation side load restoration method.

[0083] Figure 2 for Figure 1 A schematic diagram of the encoding process of the improved automatic encoder and the decoding process of the decoder in the embodiment, Figure 3 for Figure 1 The flowchart of the first layer of the improved automatic encoder in the embodiment is as follows: Figure 2 and Figure 3 As shown, further, the improved autoencoder includes three layers, and the second and third layers of the improved autoencoder both use a fully connected neural network for feature extraction; the improved autoencoder is used to encode the feature data set to obtain the encoded time series features, specifically including:

[0084] The feature dataset X is input into the first layer of the improved autoencoder, and the interdependence feature Z between the temporal variation feature quantities at each moment in the feature dataset X is analyzed through the self-attention mechanism.i Extraction; the extraction formula of the self-attention mechanism is:

[0085]

[0086] In the formula, Z i represents the interdependency feature, Q represents the query matrix, Q = XW Q , K represents the key matrix, K = XW K , V represents the value matrix, V = XW V , d k represents the dimension of the query matrix Q or key matrix K, W Q , W K , W V Represent the weight coefficients of the query matrix, key matrix, and value matrix respectively;

[0087] The gated recurrent network is used to extract the interdependency feature Z i Extract the overall time series features, including:

[0088] Reset gate calculation, the formula for reset gate calculation is: R t =σ(Z i W xr +H t-1 W hr +b r );where R t represents the output of the reset gate at time t, σ(·) represents the activation function, and Z i Represents the interdependence characteristics, H t-1 represents the hidden state at time t-1, W xr , W hr , b r Represents the learnable weight parameters of the reset gate;

[0089] Update gate calculation, the formula for update gate calculation is: Z t =σ(Z i W xz +H t-1 W hz +b z );where Z t represents the output of the update gate at time t, W xz , W hz , b z Represents the learnable weight parameters of the update gate;

[0090] Under the action of the reset gate, the candidate hidden state is calculated. The calculation formula of the candidate hidden state is: H t Y =tanh(Z i W xh +(Rt ⊙H t-1 )W hh +b h );where H t Y represents the candidate hidden state at time t, W xh , W hh , b h represents the learnable weight parameter of the candidate hidden state, ⊙ represents the Hadamard product;

[0091] Hidden state calculation, the calculation formula of the hidden state is: In the formula, H t represents the hidden state at time t;

[0092] The hidden state matrix H is constructed through the hidden state at each moment, and the hidden state matrix H is used as the extraction result of the overall time series feature;

[0093] The extraction result of the overall time series feature is used as the output result of the first layer of the improved autoencoder;

[0094] Inputting the output result of the first layer of the improved autoencoder into the second layer of the improved autoencoder;

[0095] The calculation formula for feature extraction of the second layer of the improved autoencoder is: tq =W r1 H t +b r1 Where H tq represents the output of the second layer of the improved autoencoder at time t, W r1 represents the weight coefficient of the second layer of the improved autoencoder, H t represents the time series feature at time t in the overall time series feature, that is, the hidden state at time t, b r1 represents the bias coefficient of the second layer of the improved autoencoder;

[0096] Inputting the output result of the second layer of the improved autoencoder into the third layer of the improved autoencoder;

[0097] The calculation formula for feature extraction of the third layer of the improved autoencoder is: tp =W r2 H tq +b r2 Where H tp represents the output of the third layer of the improved autoencoder at time t, W r2 represents the weight coefficient of the third layer of the improved autoencoder, b r2 represents the bias coefficient of the third layer of the improved autoencoder;

[0098] The output result of the third layer of the improved autoencoder is used as the extracted encoded temporal features.

[0099] In this application, the self-attention mechanism and gated recurrent network are used as the first layer of the improved autoencoder, referred to as the SA-GRU layer, which can achieve dimensionality reduction and feature extraction of the feature data set on the power generation side. The self-attention mechanism can realize the extraction of the interdependence relationship features between the time series change feature quantities at each moment in the feature data set, and the long-term and mutuality of the time series change feature quantities at each moment are considered and retained; the gated recurrent network can improve the memory and recursion in the feature extraction process, retain the time characteristics and overall structural characteristics of the original data, improve the efficiency of the overall time series feature extraction, and improve the prediction speed and accuracy of the comprehensive plant power consumption rate and power load.

[0100] Specifically, the decoder includes three layers, all of which use a fully connected neural network for feature reconstruction. The decoder is used to reconstruct the encoded time series features to obtain the reconstructed time series features, which specifically include:

[0101] The encoded temporal features are input into the first layer of the decoder, and the features are reconstructed through the first layer of the decoder; the feature reconstruction calculation formula in the first layer of the decoder is: A ta =W r1 ′H tp +b r1 ′; where A ta represents the output of the first layer of the decoder at time t, W r1 ′ represents the weight coefficient of the first layer of the fully connected neural network of the decoder, b r1 ′ represents the bias coefficient of the fully connected neural network of the first layer of the decoder;

[0102] The output result of the first layer of the decoder is input into the second layer of the decoder for feature reconstruction; the feature reconstruction calculation formula of the second layer of the decoder is: A tc =W r2 'A ta +b r2 ′; where A tc represents the output of the second layer of the decoder at time t, W r2 ′ represents the weight coefficient of the fully connected neural network in the second layer of the decoder, b r2 ′ represents the bias coefficient of the fully connected neural network in the second layer of the decoder;

[0103] The output result of the second layer of the decoder is input into the third layer of the decoder for feature reconstruction; the feature reconstruction calculation formula of the third layer of the decoder is: A td =W r3 'A tc +b r3′; where A td represents the output result of the third layer of the decoder at time t, that is, the reconstructed time series feature after adjustment at time t, W r3 ′ represents the weight coefficient of the fully connected neural network in the third layer of the decoder, b r3 ′ represents the bias coefficient of the fully connected neural network in the third layer of the decoder.

[0104] In this application, feature reconstruction is performed through a decoder, which improves learning efficiency and speed.

[0105] More specifically, the calculation formula for each neuron in the second layer, the third layer of the improved autoencoder and the three layers of the decoder is: i =W n x i +b n ; In the formula, y i represents the output of the i-th neuron in the output layer, x i represents the input of the i-th neuron in the input layer, W n represents the fully connected neural network weight matrix of the i-th neuron layer, b n Represents the bias matrix of the fully connected neural network in the layer where the i-th neuron is located.

[0106] More specifically, the improved autoencoder and decoder are adjusted by the loss function, specifically including: using the loss function to guide the neurons in the improved autoencoder and decoder to adjust the weight coefficients and bias coefficients of the fully connected neural network in the second layer, the third layer of the improved autoencoder, and the three layers of the decoder;

[0107] Among them, the loss function is: Where m represents the number of neurons in the output layer of the current fully connected neural network, λ represents the weight decay coefficient, l = 1, 2, ..., e, e represents the number of layers of the encoder and decoder, j = 1, 2, ..., s, s represents the number of all neurons in the current fully connected neural network, and W j l Represents the training weight value of the jth neuron in the lth layer.

[0108] In the specific implementation process, the number of layers of the encoder and the decoder are both 3, so e=3. The smaller the λ value is, the better the training effect is, so the value range of λ is between 0.0005 and 0.001.

[0109] In this application, the improved automatic encoder and decoder are adjusted through the loss function, which can ensure the consistency of the reconstructed time series features after decoding and the feature data set before encoding, making the prediction results more accurate.

[0110] Figure 4 for Figure 1 A flow chart of processing the adjusted reconstructed time series features by using a long short-term memory network in the embodiment is shown in FIG. Figure 4 As shown, further, the adjusted reconstructed time series features are processed by a long short-term memory network (LSTM network for short) to obtain the prediction results of the comprehensive plant power consumption rate and power load of the power system, specifically including:

[0111] Forget gate calculation, the calculation formula of the forget gate is: f t =sigmoid(W if A td +b if +W hf h t-1 +b hf );

[0112] In the formula, f t Represents the output of the forget gate at time t, A td represents the reconstructed timing characteristics after adjustment at time t, W if represents the weight matrix between the reconstructed time series features and the forget gate after adjustment at time t, b if represents the bias matrix between the reconstructed time series features adjusted at time t and the forget gate, W hf represents the weight matrix between the output gate and the forget gate at time t-1, b hf represents the bias matrix between the output gate and the forget gate at time t-1, h t-1 Represents the output result of the output gate at time t-1;

[0113] Input gate calculation, the calculation formula of the input gate is: c t =f t ⊙c t-1 +i t ⊙g t ;

[0114] In the formula, c t represents the output result of the input gate at time t, c t-1 represents the output result of the input gate at time t-1, i t represents the self-learning output result of the output layer feature of the input gate at time t, i t =sigmoid(W ii A td +b ii +W hi h t-1 +b hi ), W ii represents the self-learning weight matrix of the reconstructed time series features after adjustment at time t, b ii W represents the bias matrix of the reconstructed time series features after adjustment at time t; hirepresents the self-learning weight matrix of the output gate at time t-1, b hi represents the bias matrix of the output gate at time t-1, g t Indicates the output result of the input gate at time t based on the learning of the assigned weight matrix features, g t =tanh(W ig A td +b ig +W hg h t-1 +b hg ), W ig represents the weight matrix between the reconstructed timing features and the input gate after adjustment at time t, b ig represents the bias matrix between the reconstructed timing features and the input gate after adjustment at time t, W hg represents the weight matrix between the output gate and the input gate at time t-1, b hg Represents the bias matrix between the output gate and the input gate at time t-1;

[0115] Output gate calculation, the calculation formula of the output gate is: h t =o t ⊙tanh(c t );

[0116] In the formula, h t represents the output result of the output gate at time t, o t =sigmoid(W io A td +b io +W ho h t-1 +b ho ), W io represents the weight matrix between the reconstructed timing features and the output gate after adjustment at time t, b io represents the bias matrix between the reconstructed timing features and the output gate after adjustment at time t, W ho represents the weight matrix between the output gate at time t-1 and the output gate at time t, b ho Represents the bias matrix between the output gate at time t-1 and the output gate at time t;

[0117] The comprehensive plant power consumption rate η at time t is calculated by the first fully connected neural network t , η t =W o h t +b o Where W o represents the weight matrix of the first fully connected neural network, b o represents the bias matrix of the first fully connected neural network;

[0118] The power load forecast result L after the power generation side is restored at time t is calculated by the comprehensive power consumption rate at time t t ′,L t ′=G t +S t +W t +F t +P t +D t (1-η t ).

[0119] In this application, the long short-term memory network and the first fully connected neural network are used to predict the comprehensive plant power consumption rate, which effectively models and learns nonlinear and long-sequence data, improves the generalization ability of the model, and improves adaptability. The long short-term memory network can perform well in the task of modeling data with long-term dependencies, and is suitable for use in complex output scenarios involving thermal power units, hydropower units, wind power units, gas units, centralized photovoltaic power stations, distributed photovoltaics, etc. in this application. Before the long short-term memory network processes the adjusted reconstructed timing features, the self-attention mechanism and the gated recurrent network are used to reduce the dimension and extract the overall timing features, which solves the problem that the long short-term memory network needs to train and learn a large amount of sample data for a long time, which is more efficient and accurate. The calculated power load forecast result L after restoration on the power generation side at time t t ′, can be directly compared with the planned load L of the whole area at time t t The comparison can also be used for error analysis and accuracy improvement of load forecasting.

[0120] In the specific implementation process, the historical comprehensive plant power consumption rate can be calculated based on the accurate historical output and historical load data of each unit, so as to generate the output label of the deep learning model for supervised training and learning. It is suitable for the calculation of historical label data, ensuring the accuracy of historical label data and improving prediction accuracy.

[0121] Among them, the calculation formula for the historical comprehensive plant power consumption rate at time r is:

[0122]

[0123] In the formula, G r " represents the historical output of the thermal power unit at time r, S r " represents the historical output of the hydropower unit at time r, W r " represents the historical output of the wind turbine at time r, F r " represents the historical output of the gas unit at time r, P r " represents the historical output of the centralized photovoltaic power station at time r, D r ” represents the historical output of distributed photovoltaic power generation at time r, L r” represents the historical load of the whole area at time r;

[0124] The output label sequence of the historical comprehensive power consumption rate of the plant at R moments is constructed as follows:

[0125] Y=[η1”,……,η r ”,……,η R ”], where r = 1, 2, …, R, and R represents the number of historical moments.

[0126] The present application also provides a deep learning-based integrated power system plant power consumption prediction system, including a module for implementing a deep learning-based integrated power system plant power consumption prediction method as described in any of the above items.

[0127] Figure 5 A functional structure diagram of a power system comprehensive plant power consumption prediction system based on deep learning provided in an embodiment of the present application, such as Figure 5 Specifically, a power system comprehensive plant power consumption prediction system based on deep learning includes:

[0128] The collection module 10 is used to collect characteristic data at each time in the forecast area; wherein the characteristic data includes the planned output of each unit and the planned load of the whole area;

[0129] A feature data set construction module 20, used to construct a feature data set according to the feature data at each moment;

[0130] An encoding module 30, used to encode the feature data set using an improved automatic encoder to obtain encoded time series features;

[0131] A reconstruction module 40, configured to reconstruct the encoded timing characteristics using a decoder to obtain a reconstructed timing characteristics;

[0132] A judgment module 50 is used to judge whether the reconstructed time series feature is consistent with the feature data set;

[0133] An adjustment module 60, used for adjusting the improved automatic encoder and decoder through a loss function when the reconstructed time series feature is inconsistent with the feature data set, until the adjusted reconstructed time series feature is consistent with the feature data set;

[0134] The prediction module 70 is used to process the adjusted reconstructed time series characteristics through a long short-term memory network to obtain the prediction results of the comprehensive plant power consumption rate and power load of the power system.

[0135] Figure 6 for Figure 5 The functional structure diagram of the feature data set construction module in the embodiment is as follows: Figure 6 As shown, specifically, the feature data set construction module 20 includes:

[0136] The time feature vector construction unit 201 is used to construct the corresponding time feature vector according to the collected feature data at each time in the prediction area; the feature vector x at time t t for:

[0137] x t =[G t ,S t ,W t ,F t ,P t ,D t ,L t ];

[0138] In the formula, G t represents the planned output of the thermal power unit at time t, S t represents the planned output of the hydropower unit at time t, W t represents the planned output of the wind turbine at time t, F t represents the planned output of the gas unit at time t, P t represents the planned output of the centralized photovoltaic power station at time t, D t represents the planned output of distributed photovoltaic power generation at time t, L t represents the planned load of the whole area at time t;

[0139] The time series variation feature quantity construction unit 202 is used to construct the time series variation feature quantity X at time t. t For: X t =[x t-1 x t x t+1 ] T ; In the formula, x t-1 represents the eigenvector at time t-1, x t+1 Represents the feature vector at time t+1;

[0140] The feature data set construction unit 203 is used to construct a feature data set: X = [X1, ..., X t ,……,X N ], where X N It represents the time series variation feature at N moments, t=1, 2, ..., N, where N represents the number of acquisitions, i.e. the number of moments.

[0141] Figure 7 for Figure 5 The functional structure diagram of the encoding module in the embodiment is as follows: Figure 7 As shown, specifically, the encoding module 30 includes:

[0142] The self-attention mechanism extraction unit 301 is used to input the feature data set X into the first layer of the improved autoencoder, and use the self-attention mechanism to extract the interdependence relationship feature Z between the time series change feature quantities at each moment in the feature data set X. i Extraction; the extraction formula of the self-attention mechanism is:

[0143]

[0144] In the formula, Z i represents the interdependency feature, Q represents the query matrix, Q = XW Q , K represents the key matrix, K = XW K , V represents the value matrix, V = XW V , d k represents the dimension of the query matrix Q or key matrix K, W Q , W K , W V Represent the weight coefficients of the query matrix, key matrix, and value matrix respectively;

[0145] The gated recurrent network extraction unit 302 is used to extract the interdependence feature Z using the gated recurrent network. i Extract the overall time series features, including:

[0146] The reset gate calculation unit 3021 is used to reset the gate calculation. The formula for resetting the gate calculation is: R t =σ(Z i W xr +H t-1 W hr +b r );where R t represents the output of the reset gate at time t, σ(·) represents the activation function, and Z i Represents the interdependence characteristics, H t-1 represents the hidden state at time t-1, W xr , W hr , b r Represents the learnable weight parameters of the reset gate;

[0147] The update gate calculation unit 3022 is used for updating the gate calculation. The formula for updating the gate calculation is: t =σ(Z i W xz +H t-1 W hz +b z );where Z t represents the output of the update gate at time t, W xz , W hz , b zRepresents the learnable weight parameters of the update gate;

[0148] The candidate hidden state calculation unit 3023 is used to calculate the candidate hidden state under the action of the reset gate. The calculation formula of the candidate hidden state is: t Y =tanh(Z i W xh +(R t ⊙H t-1 )W hh +b h );where H t Y represents the candidate hidden state at time t, W xh , W hh , b h represents the learnable weight parameter of the candidate hidden state, ⊙ represents the Hadamard product;

[0149] Hidden state calculation unit 3024 is used for hidden state calculation. The calculation formula of the hidden state is: In the formula, H t represents the hidden state at time t;

[0150] A hidden state matrix construction unit 3025 is used to construct a hidden state matrix H through the hidden states at each moment, and use the hidden state matrix H as the extraction result of the overall time series feature;

[0151] A first processing unit 303 is used to use the extraction result of the overall time series feature as the output result of the first layer of the improved automatic encoder;

[0152] A second processing unit 304 is used to input the output result of the first layer of the improved auto-encoder into the second layer of the improved auto-encoder;

[0153] The calculation formula for feature extraction of the second layer of the improved autoencoder is: tq =W r1 H t +b r1 Where H tq represents the output of the second layer of the improved autoencoder at time t, W r1 represents the weight coefficient of the second layer of the improved autoencoder, H t represents the time series feature at time t in the overall time series feature, that is, the hidden state at time t, b r1 represents the bias coefficient of the second layer of the improved autoencoder;

[0154] A third processing unit 305 is used to input the output result of the second layer of the improved autoencoder into the third layer of the improved autoencoder;

[0155] The calculation formula for feature extraction of the third layer of the improved autoencoder is: tp =W r2 H tq +b r2 Where H tp represents the output of the third layer of the improved autoencoder at time t, W r2 represents the weight coefficient of the third layer of the improved autoencoder, b r2 represents the bias coefficient of the third layer of the improved autoencoder;

[0156] The coding time series feature extraction unit 306 is used to use the output result of the third layer of the improved automatic encoder as the extracted coding time series feature.

[0157] Figure 8 for Figure 5 The functional structure diagram of the reconstruction module in the embodiment is as follows: Figure 8 As shown, specifically, the reconstruction module 40 includes:

[0158] The first reconstruction unit 401 is used to input the encoded time series features into the first layer of the decoder and perform feature reconstruction through the first layer of the decoder; the feature reconstruction calculation formula in the first layer of the decoder is: ta =W r1 ′H tp +b r1 ′; where A ta represents the output of the first layer of the decoder at time t, W r1 ′ represents the weight coefficient of the first layer of the fully connected neural network of the decoder, b r1 ′ represents the bias coefficient of the fully connected neural network of the first layer of the decoder;

[0159] The second reconstruction unit 402 is used to input the output result of the first layer of the decoder into the second layer of the decoder for feature reconstruction; the feature reconstruction calculation formula of the second layer of the decoder is: tc =W r2 'A ta +b r2 ′; where A tc represents the output of the second layer of the decoder at time t, W r2 ′ represents the weight coefficient of the fully connected neural network in the second layer of the decoder, b r2 ′ represents the bias coefficient of the fully connected neural network in the second layer of the decoder;

[0160] The third reconstruction unit 403 is used to input the output result of the second layer of the decoder into the third layer of the decoder for feature reconstruction; the feature reconstruction calculation formula of the third layer of the decoder is: td =W r3'A tc +b r3 ′; where A td represents the output result of the third layer of the decoder at time t, that is, the reconstructed time series feature after adjustment at time t, W r3 ′ represents the weight coefficient of the fully connected neural network in the third layer of the decoder, b r3 ′ represents the bias coefficient of the fully connected neural network in the third layer of the decoder.

[0161] Fig. 9 for Figure 5 The functional structure diagram of the prediction module in the embodiment is as follows: Fig. 9 As shown, specifically, the prediction module 70 includes:

[0162] The forget gate calculation unit 701 is used for forget gate calculation. The calculation formula of the forget gate is: t =sigmoid(W if A td +b if +W hf h t-1 +b hf );

[0163] In the formula, f t Represents the output of the forget gate at time t, A td represents the reconstructed timing characteristics after adjustment at time t, W if represents the weight matrix between the reconstructed time series features and the forget gate after adjustment at time t, b if represents the bias matrix between the reconstructed time series features adjusted at time t and the forget gate, W hf represents the weight matrix between the output gate and the forget gate at time t-1, b hf represents the bias matrix between the output gate and the forget gate at time t-1, h t-1 Represents the output result of the output gate at time t-1;

[0164] The input gate calculation unit 702 is used for input gate calculation. The calculation formula of the input gate is: t =f t ⊙c t-1 +i t ⊙g t ;

[0165] In the formula, c t represents the output result of the input gate at time t, c t-1 represents the output result of the input gate at time t-1, i t represents the self-learning output result of the output layer feature of the input gate at time t, i t =sigmoid(W ii A td+b ii +W hi h t-1 +b hi ), W ii represents the self-learning weight matrix of the reconstructed time series features after adjustment at time t, b ii W represents the bias matrix of the reconstructed time series features after adjustment at time t; hi represents the self-learning weight matrix of the output gate at time t-1, b hi represents the bias matrix of the output gate at time t-1, g t Indicates the output result of the input gate at time t based on the learning of the assigned weight matrix features, g t =tanh(W ig A td +b ig +W hg h t-1 +b hg ), W ig represents the weight matrix between the reconstructed timing features and the input gate after adjustment at time t, b ig represents the bias matrix between the reconstructed timing features and the input gate after adjustment at time t, W hg represents the weight matrix between the output gate and the input gate at time t-1, b hg Represents the bias matrix between the output gate and the input gate at time t-1;

[0166] The output gate calculation unit 703 is used for output gate calculation. The calculation formula of the output gate is: t =o t ⊙tanh(c t );

[0167] In the formula, h t represents the output result of the output gate at time t, o t =sigmoid(W io A td +b io +W ho h t-1 +b ho ), W io represents the weight matrix between the reconstructed timing features and the output gate after adjustment at time t, b io represents the bias matrix between the reconstructed timing features and the output gate after adjustment at time t, W ho represents the weight matrix between the output gate at time t-1 and the output gate at time t, b ho Represents the bias matrix between the output gate at time t-1 and the output gate at time t;

[0168] The comprehensive plant power consumption rate calculation unit 704 is used to calculate the comprehensive plant power consumption rate η at time t through the first fully connected neural networkt , η t =W o h t +b o Where W o represents the weight matrix of the first fully connected neural network, b o represents the bias matrix of the first fully connected neural network;

[0169] The power load calculation unit 705 is used to calculate the power load forecast result L after the power generation side is restored at time t through the comprehensive plant power consumption rate at time t. t ′,L t ′=G t +S t +W t +F t +P t +D t (1-η t ).

[0170] The present application also provides an electronic device, including:

[0171] Memory;

[0172] Processor; and

[0173] Computer programs;

[0174] Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the deep learning-based comprehensive plant power consumption prediction method for the power system as described in any one of the above.

[0175] The present application also provides a computer-readable storage medium having a computer program stored thereon; the computer program is executed by a processor to implement a method for predicting comprehensive plant power consumption of a power system based on deep learning as described in any one of the above items.

[0176] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiment, complete software embodiment, or the embodiment in combination with software and hardware. Moreover, the present application can adopt the form of the computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be realized in various computer languages, for example, C language, VHDL language, Verilog language, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0177] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes 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 generate 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.

[0178] 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.

[0179] 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.

[0180] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0181] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0182] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for predicting comprehensive power consumption of power system based on deep learning, characterized in that: include: Collect characteristic data at each moment in the forecast area, and construct a characteristic data set according to the characteristic data at each moment; wherein the characteristic data includes the planned output of each unit and the planned load of the whole area; the characteristic data set is constructed according to the characteristic data at each moment, specifically including: The corresponding time feature vector is constructed based on the characteristic data collected at each time in the prediction area; the feature vector x at time t t for: x t =[G t ,S t ,W t ,F t ,P t ,D t ,L t ]; In the formula, G t represents the planned output of the thermal power unit at time t, S t represents the planned output of the hydropower unit at time t, W t represents the planned output of the wind turbine at time t, F t represents the planned output of the gas unit at time t, P t represents the planned output of the centralized photovoltaic power station at time t, D t represents the planned output of distributed photovoltaic power generation at time t, L t represents the planned load of the whole area at time t; Construct the time series variation feature X at time t t For: X t =[x t-1 x t x t+1 ] T ; In the formula, x t-1 represents the eigenvector at time t-1, x t+1 Represents the feature vector at time t+1; Construct feature data set: X = [X1, ..., X t ,……,X N ], where X N Represents the time series variation feature at N moments, t = 1, 2, ..., N, N represents the number of acquisitions, i.e., the number of moments; An improved automatic encoder is used to encode the feature data set to obtain encoded time series features, and then a decoder is used to reconstruct the encoded time series features to obtain reconstructed time series features; Determine whether the reconstructed time series features are consistent with the feature data set. If not, adjust the improved automatic encoder and decoder through the loss function until the adjusted reconstructed time series features are consistent with the feature data set. The adjusted reconstructed time series features are processed through the long short-term memory network to obtain the prediction results of the comprehensive power consumption rate and power load of the power system, including: Forget gate calculation, the calculation formula of the forget gate is: f t =sigmoid(W if A td +b if +W hf h t-1 +b hf ); In the formula, f t Represents the output of the forget gate at time t, A td represents the reconstructed timing characteristics after adjustment at time t, W if represents the weight matrix between the reconstructed time series features and the forget gate after adjustment at time t, b if represents the bias matrix between the reconstructed time series features adjusted at time t and the forget gate, W hf represents the weight matrix between the output gate and the forget gate at time t-1, b hf represents the bias matrix between the output gate and the forget gate at time t-1, h t-1 Represents the output result of the output gate at time t-1; Input gate calculation, the calculation formula of the input gate is: c t =f t ⊙c t-1 +i t ⊙g t ; In the formula, c t represents the output result of the input gate at time t, c t-1 represents the output result of the input gate at time t-1, i t represents the self-learning output result of the output layer feature of the input gate at time t, i t =sigmoid(W ii A td +b ii +W hi h t-1 +b hi ), W ii represents the self-learning weight matrix of the reconstructed time series features after adjustment at time t, b ii W represents the bias matrix of the reconstructed time series features after adjustment at time t; hi represents the self-learning weight matrix of the output gate at time t-1, b hi represents the bias matrix of the output gate at time t-1, g t Indicates the output result of the input gate at time t based on the learning of the assigned weight matrix features, g t =tanh(W ig A td +b ig +W hg h t-1 +b hg ), W ig represents the weight matrix between the reconstructed timing features and the input gate after adjustment at time t, b ig represents the bias matrix between the reconstructed timing features and the input gate after adjustment at time t, W hg represents the weight matrix between the output gate and the input gate at time t-1, b hg Represents the bias matrix between the output gate and the input gate at time t-1; Output gate calculation, the calculation formula of the output gate is: h t =o t ⊙tanh(c t ); In the formula, h t represents the output result of the output gate at time t, o t =sigmoid(W io A td +b io +W ho h t-1 +b ho ), W io represents the weight matrix between the reconstructed timing features and the output gate after adjustment at time t, b io represents the bias matrix between the reconstructed timing features and the output gate after adjustment at time t, W ho represents the weight matrix between the output gate at time t-1 and the output gate at time t, b ho Represents the bias matrix between the output gate at time t-1 and the output gate at time t; The comprehensive plant power consumption rate η at time t is calculated by the first fully connected neural network t , η t =W o h t +b o Where W o represents the weight matrix of the first fully connected neural network, b o represents the bias matrix of the first fully connected neural network; The power load forecast result L after the power generation side is restored at time t is calculated by the comprehensive power consumption rate at time t t ′,L t ′=G t +S t +W t +F t +P t +D t (1-η t ).

2. The method for predicting power consumption of a power system based on deep learning according to claim 1 is characterized in that: The improved autoencoder comprises three layers; The improved automatic encoder is used to encode the feature data set to obtain the encoded time series features, which specifically includes: The feature data set is input into the first layer of the improved autoencoder, and the interdependence features between the time series change feature quantities at each moment in the feature data set are extracted through the self-attention mechanism; the gated recurrent network is used to extract the overall time series features from the extracted interdependence features; the extraction result of the overall time series features is used as the output result of the first layer of the improved autoencoder; Inputting the output result of the first layer of the improved autoencoder into the second layer of the improved autoencoder; Inputting the output result of the second layer of the improved autoencoder into the third layer of the improved autoencoder; The output result of the third layer of the improved autoencoder is used as the extracted encoded temporal features.

3. The method for predicting power consumption of a power system based on deep learning according to claim 2 is characterized in that: The extraction formula of the self-attention mechanism is: In the formula, Z i represents the interdependency feature, Q represents the query matrix, Q = XW Q , K represents the key matrix, K = XW K , V represents the value matrix, V = XW V , d k represents the dimension of the query matrix Q or key matrix K, W Q , W K , W V Represent the weight coefficients of the query matrix, key matrix, and value matrix respectively.

4. The method for predicting power consumption of a power system based on deep learning according to claim 2 is characterized in that: The extracting of the overall temporal features from the extracted interdependence features using a gated recurrent network specifically includes: Reset gate calculation, the formula for reset gate calculation is: R t =σ(Z i W xr +H t-1 W hr +b r ); In the formula, R t represents the output of the reset gate at time t, σ(·) represents the activation function, and Z i Represents the interdependence characteristics, H t-1 represents the hidden state at time t-1, W xr , W hr , b r Represents the learnable weight parameters of the reset gate; Update gate calculation, the formula for update gate calculation is: Z t =σ(Z i W xz +H t-1 W hz +b z ); In the formula, Z t represents the output of the update gate at time t, W xz , W hz , b z Represents the learnable weight parameters of the update gate; Under the action of the reset gate, the candidate hidden state is calculated. The calculation formula of the candidate hidden state is: H t Y =tanh(Z i W xh +(R t ⊙H t-1 )W hh +b h ); In the formula, H t Y represents the candidate hidden state at time t, W xh , W hh , b h represents the learnable weight parameter of the candidate hidden state, ⊙ represents the Hadamard product; Hidden state calculation, the calculation formula of the hidden state is: In the formula, H t represents the hidden state at time t; The hidden state matrix is ​​constructed through the hidden state at each moment, and the hidden state matrix is ​​used as the extraction result of the overall time series feature.

5. The method for predicting power consumption of a power system based on deep learning according to claim 2 is characterized in that: The second and third layers of the improved autoencoder both use a fully connected neural network for feature extraction; The calculation formula for feature extraction of the second layer of the improved autoencoder is: tq =W r1 H t +b r1 Where H tq represents the output of the second layer of the improved autoencoder at time t, W r1 represents the weight coefficient of the second layer of the improved autoencoder, H t represents the time series feature at time t in the overall time series feature, that is, the hidden state at time t, b r1 represents the bias coefficient of the second layer of the improved autoencoder; The calculation formula for feature extraction of the third layer of the improved autoencoder is: tp =W r2 H tq +b r2 Where H tp represents the output of the third layer of the improved autoencoder at time t, W r2 represents the weight coefficient of the third layer of the improved autoencoder, b r2 Represents the bias coefficients of the third layer of the improved autoencoder.

6. A power system comprehensive plant power consumption prediction system based on deep learning, characterized in that: It includes a module for implementing the deep learning-based comprehensive plant power consumption prediction method for a power system as described in any one of claims 1-5.

7. An electronic device, characterized in that: include: Memory; processor; as well as Computer programs; Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the deep learning-based comprehensive plant power consumption prediction method for a power system as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: A computer program is stored thereon; the computer program is executed by a processor to implement a method for predicting comprehensive plant power consumption of a power system based on deep learning as described in any one of claims 1-5.

Images