Big data driven intelligent energy management system and method

Abstract

The invention discloses a big data driven intelligent energy management system and method, and belongs to the technical field of energy management. The system comprises: a feature extraction module for analyzing a Granger causal relationship and time sequence features among electrical equipment parameters through a first deep learning model, generating an association weight matrix and constructing a standardized data stream; the intelligent prediction module is used for fusing the standardized data stream and external environment data, inputting the data into a second deep learning model after time alignment and normalization, and synchronously predicting the power demand and cost; and the strategy optimization module is used for designing a reinforcement learning framework based on the prediction result, and generating a control strategy by taking the minimum cost, the equipment scheduling frequency and the charging and discharging fluctuation as targets. According to the method, the characteristic interpretability is improved through causal analysis, the prediction precision is enhanced through a bidirectional sequential network, a dynamic optimization closed loop is realized by adopting reinforcement learning, and the intelligent level of energy management and the resource utilization efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention relates to a field of intelligent energy management and management system driven by big data. Background Art

[0002] With the continuous growth of global energy demand and the rapid development of smart grid technology, efficient and accurate energy management has become a core requirement for achieving a low-carbon and intelligent energy system.

[0003] The core bottlenecks of existing technologies are: ① The lack of an interpretable framework for mining the associated features of power equipment parameters makes it difficult to reveal physical coupling relationships through causal testing; ② The imperfect spatiotemporal alignment and feature fusion mechanisms of multi-source data limit the generalization ability of the prediction model; and ③ The lack of a dynamic optimization closed loop in the energy control strategy makes it impossible to adaptively adjust the equipment operating status based on real-time prediction results. The above problems directly lead to insufficient prediction accuracy and lagging control strategies in the energy management system, which restricts the improvement of energy utilization efficiency and the optimization of overall costs. Therefore, there is an urgent need to build an intelligent energy management system that integrates big data analysis, deep learning, and reinforcement learning to achieve full-process intelligence from data feature analysis to strategy generation. Summary of the Invention

[0004] In order to solve the technical problems mentioned in the current background technology, the present invention proposes a big data driven intelligent energy management system and method.

[0005] To this end, the technical solution adopted in the present invention is as follows:

[0006] A big data-driven intelligent energy management system, characterized in that the system includes:

[0007] A feature extraction module analyzes the power equipment parameters using the trained first deep learning model to generate correlation weights; and combines the correlation weights with the power equipment parameters to generate a standardized data stream.

[0008] An intelligent prediction module sequentially performs time alignment, normalization, and splicing operations on the standardized data stream and the external environment data stream to generate a spliced ​​vector; divides the spliced ​​vector into second samples to generate a power prediction data set; and trains a second deep learning model using the power prediction data set. When the second deep learning model converges, it is used to predict power demand and power cost.

[0009] A strategy optimization module receives the predicted power demand and power cost and the external environment data stream, and generates an energy control strategy in combination with a reinforcement learning algorithm

[0010] Furthermore, the power equipment parameters are expressed in the form of parameter time series, which is as follows:

[0011] X t =[X t,1 ,X t,2 ,...,X t,I ]

[0012] where X t,i is the parameter of the ith power equipment at time t; I is the number of power equipment parameters;

[0013] The parameter time series of the historical power equipment parameters is cut into N first samples to generate a power equipment parameter data set, each first sample has a length of p, and the first sample is represented by [X t-p ,X t-p+1 ,...,X t-1 ];

[0014] The training process of the first deep learning model is as follows:

[0015] The first layer is the input layer: it receives the first sample;

[0016] The second layer is the time series analysis layer: the first sample is input into the LSTM unit, and the time series feature vector is obtained by combining the global average pooling operation;

[0017] The third layer is the parameter association analysis layer: the first sample is simultaneously input into the complete regression model and the residual regression model to predict the power equipment parameters at the next moment;

[0018] Define parameter relationship matrix M={M ij}, matrix element M ij Represents the relationship coefficient between the i-th power equipment parameter and the j-th power equipment parameter;

[0019] For the matrix element M ij The formula of the established complete regression model is as follows:

[0020]

[0021] in, In the complete regression model, the jth power equipment parameter at time t predicted by the ith power equipment parameter is X t,-i represents the parameters of other power equipment except the parameters of the i-th power equipment at time t; α j is the intercept term; β jk and γ jk is the regression coefficient;∈ t is the random error term;

[0022] For the matrix element M ij The formula of the residual regression model established is as follows:

[0023]

[0024] in, In the residual regression model, it represents the jth power equipment parameter at time t predicted by the i-th power equipment parameter;

[0025] Calculate the F statistic using the formula:

[0026]

[0027] When F>F α (p, N-2p-1), it is considered that the parameter i of the i-th power equipment has a Granger causal relationship with the parameter j of the j-th power equipment, and the M ij =β jk ,

[0028] Among them, F α (p,N-2p-1) represents the quantile when the cumulative probability on the right is α in the F distribution with (p,N-2p-1) degrees of freedom;

[0029] in,

[0030] The fourth layer is the output layer: the attention mechanism is used to calculate the temporal feature score and parameter relationship score respectively. The formulas are as follows:

[0031] a trend =Softmax(W trend T+b trend )

[0032] a rel =Softmax(W rel M+b rel )

[0033] Among them, W trend 、W rel is the weight coefficient, b trend and b rel is the bias term;

[0034] The temporal feature score and the parameter relationship score are combined to obtain the association weight Ω, which is as follows:

[0035] Ω=a trend ⊙T+a rel ⊙M

[0036] Where, ⊙term-by-term product operator = {Ω ij}, where Ω ij represents the association weight between power equipment parameter i and power equipment parameter j;

[0037] The dimension of the time feature vector is the same as the dimension of the parameter association matrix, so the number of LSTM units is equal to the square of the number of power equipment parameters.

[0038] Furthermore, the loss function of the first deep learning model is:

[0039]

[0040] in, is the weighted prediction value,

[0041] After the first deep learning model training converges, the optimal association weight is used; the optimal association weight is multiplied by the transposed matrix of the power equipment parameter to generate a standardized data stream The formula is as follows:

[0042]

[0043] in, Represents X t The transposed matrix of .

[0044] Furthermore, the formula for the splicing operation is as follows:

[0045]

[0046] Among them, Y t represents the splicing vector, S t For external environment data flow;

[0047] The external environment data flow is represented as:

[0048] S t =[S t, 1,S t, 2,...,S t,s ]

[0049] Among them S t,s is the sth external environment data at time t; s is the number of external environment data.

[0050] Furthermore, the second sample length is q, and the second sample represents [Y t-q ,Y t-q+1 ,...,Y t-1 ];

[0051] The second deep learning model takes the second sample as input, and the training process is as follows:

[0052] The first layer is the input layer: receiving the second sample;

[0053] The second layer is an embedding layer: the second sample is linearly transformed into an embedding vector, and the length of the embedding vector is equal to the length of the second sample;

[0054] The third layer is a Bi-GRU layer: it is used to extract the bidirectional time series features of the embedding vector to generate a bidirectional time series feature vector.

[0055] The fourth layer is the output layer: it contains the power demand forecast branch and the power cost forecast branch, which are used to output the power demand and power cost of a fixed time period in the future respectively;

[0056] The power demand forecasting branch adopts linear transformation forecasting, and the power cost forecasting branch adopts nonlinear transformation forecasting, and the formulas are respectively expressed as follows:

[0057]

[0058] in, represents the power demand in the future H time period, represents the electricity cost in the future H time period; W D 、W C1 and W C2 is the coefficient weight, b D 、b C1 and b C2 is the bias weight.

[0059] Furthermore, the loss function of the second deep learning prediction model includes a demand loss term and a cost loss term. The formula for the demand loss term is as follows:

[0060]

[0061] in, represents the predicted power demand at time t+h of the g-th second sample, represents the actual power demand at time t+h of the g-th second sample; represents the predicted power demand at time t+h of the g-th second sample, represents the actual power demand at time t+h of the g-th second sample; G represents the number of the second samples;

[0062] The loss function of the second deep learning model is expressed as: L1=L D +λL C , λ is the balance coefficient.

[0063] Furthermore, the training process of the reinforcement learning is as follows:

[0064] Define the state space as:

[0065]

[0066] Among them, S t,1 represents the electricity price at time t; represents the power demand in the future H time period, C t:t+H-1 represents the electricity cost in the future H time period;

[0067] The action space of the energy control strategy is defined as:

[0068] a t ={U t ,E t}

[0069] Among them, U t represents the power equipment dispatch vector at time t, E t Indicates the energy storage charging and discharging power of the power equipment at time t;

[0070] Define a reward function, which is to minimize the difference between the electricity cost and the actual electricity cost, minimize the number of power equipment dispatches, and minimize the number of power equipment dispatches. There are power cost reward items, power equipment dispatch reward items, and power equipment energy storage charging and discharging reward items, respectively. The formula is as follows:

[0071]

[0072] The state z at time t-1 t-1 Input to the Actor decision network to generate action a at time t t , and collect the state z at time t t , change action a t and state z t Input into the Critic value network to judge action a t The value of , the Critic value network includes the reward function;

[0073] Optimize the Critic value network through the mean square error function until convergence to maximize the reward; optimize the Actor decision network through the policy gradient algorithm until convergence;

[0074] After the Critic value network and the Actor decision network converge, the optimal reinforcement learning algorithm is obtained.

[0075] Furthermore, the optimal reinforcement learning algorithm receives the predicted power demand and power cost and the external environment data stream to generate actions of the energy control strategy.

[0076] The present invention also proposes a big data driven intelligent energy management method, characterized in that the method comprises the following steps:

[0077] Analyzing the power equipment parameters using the trained first deep learning model to generate correlation weights; and combining the correlation weights and the power equipment parameters to generate a standardized data stream;

[0078] The standardized data stream and the external environment data stream are sequentially time-aligned, normalized, and spliced ​​to generate a spliced ​​vector; the spliced ​​vector is divided into a second sample to generate a power prediction data set; the power prediction data set is used to train a second deep learning model, and when the second deep learning model converges, the model is used to predict power demand and power cost;

[0079] The predicted power demand and power cost and the external environment data stream are received, and an energy control strategy is generated in combination with a reinforcement learning algorithm.

[0080] Compared with the prior art, the advantages of the present invention are:

[0081] 1. This paper analyzes the physical correlation between power equipment parameters through causal testing and attention mechanism, combines the timing analysis network to capture operating trends and periodic patterns, generates an interpretable dynamic weight matrix, and realizes in-depth mining of parameter correlation features.

[0082] 2. This invention integrates power equipment parameters and external environmental data to construct a multidimensional input feature space. It then extracts bidirectional dependencies from historical data using a bidirectional time series network. It then employs two task prediction branches to simultaneously output power demand and cost, providing comprehensive and reliable forward-looking information for energy planning.

[0083] 3. This invention leverages a reinforcement learning model to dynamically optimize equipment scheduling and energy storage strategies to balance costs. Combined with a closed-loop optimization mechanism that provides real-time feedback, the system automatically adapts to fluctuations in energy supply and demand, enabling intelligent management of the entire process from prediction to control. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0085] Figure 1 Schematic diagram of the feature extraction module of the present invention;

[0086] Figure 2 This is a working diagram of the intelligent prediction module of the present invention;

[0087] Figure 3 This is a working diagram of the strategy optimization module of the present invention. DETAILED DESCRIPTION

[0088] To achieve the above objectives, the present invention is implemented through the following technical solutions. The present invention provides a big data-driven intelligent energy management system, which includes:

[0089] M1, feature extraction module, uses the first deep learning model to analyze the power equipment parameters to generate correlation weights; combines the correlation weights and power equipment parameters to generate a standardized data stream; see Figure 1 .

[0090] In this embodiment, the power equipment parameters are expressed in the form of parameter time series, and the parameter time series is expressed as: t =[X t,1 ,X t,2 ,...,X t,I ], where X t,i is the parameter of the i-th power equipment at time t; I is the number of power equipment parameters.

[0091] In this embodiment, the parameter time series of the historical power equipment parameters is cut into N first samples to generate a power equipment parameter data set. The length of each first sample is p, and the first sample is represented by [X t-p ,X t-p+1 ,...,X t-1 ]. The first deep learning model is trained based on the power equipment parameter data set.

[0092] In this embodiment, the first deep learning model takes the first sample as input and outputs the associated weight;

[0093] The training process of the first deep learning model is as follows:

[0094] The first layer is the input layer: it receives the first sample.

[0095] The second layer is the time series analysis layer: the first sample is input into the LSTM unit to obtain a time series feature vector; the LSTM unit consists of an input gate, a forget gate, and an output gate. The final output state is a hidden state vector. The detailed structure of LSTM is not repeated here.

[0096] A global average pooling operation is used to calculate the average value of the hidden state features at all times to obtain the temporal feature vector T.

[0097] The third layer is the parameter association analysis layer: the first sample is input into the complete regression model and the residual regression model at the same time to predict the power equipment parameters at the next moment; at the same time, the parameter relationship matrix M = {M ij}, matrix element M ijRepresents the relationship coefficient between the i-th power equipment parameter and the j-th power equipment parameter.

[0098] For the matrix element M ij The formula of the established complete regression model is as follows:

[0099]

[0100] in, In the complete regression model, the jth power equipment parameter at time t predicted by the ith power equipment parameter is X t,-i represents the parameters of other power equipment except the parameters of the i-th power equipment at time t; α j is the intercept term; β jk and γ jk is the regression coefficient;∈ t is the random error term.

[0101] For the matrix element M ij The formula of the residual regression model established is as follows:

[0102]

[0103] in, In the residual regression model, it represents the jth power equipment parameter at time t predicted by the i-th power equipment parameter;

[0104] Calculate the F statistic using the formula:

[0105]

[0106] When F>F α (p, N-2p-1), it is considered that the parameter i of the i-th power equipment has a Granger causal relationship with the parameter j of the j-th power equipment, and the M ij =β jk ,

[0107] Among them, F α (p,N-2p-1) represents the quantile when the cumulative probability on the right side is α in an F-distribution with (p,N-2p-1) degrees of freedom.

[0108] in,

[0109] The fourth layer is the output layer: the attention mechanism is used to calculate the temporal feature score and parameter relationship score respectively. The formulas are as follows:

[0110] a trend =Softmax(W trend T+b trend )

[0111] a rel =Softmax(W rel M+b rel )

[0112] Among them, Softmax() is the Softmax function, W trend 、W rel is the weight coefficient, b trend and b rel is the bias term.

[0113] The association weight parameter Ω is obtained by fusing the time feature score and the parameter relationship score. The formula is as follows:

[0114] Ω=a trend ⊙T+a rel ⊙M

[0115] Where, ⊙term-by-term product operator = {Ω ij}, where Ω ij Represents the association weight between power equipment parameter i and power equipment parameter j.

[0116] In this embodiment, the dimension of the time feature vector is the same as the dimension of the parameter association matrix, so it is necessary to ensure that the number of LSTM units is equal to the square of the number of power equipment parameters, that is, I 2 .

[0117] The loss function of the first deep learning model is:

[0118]

[0119] in, is the weighted prediction value,

[0120] The coupling relationship between the power equipment parameters is optimized through the loss function.

[0121] In some embodiments, the power device parameters include current, voltage, and temperature.

[0122] After the first deep learning model training converges, the optimal association weight is obtained. The optimal association weight is multiplied by the transposed matrix of the power equipment parameter to generate a standardized data stream The formula is as follows:

[0123]

[0124] in, Represents X t The transposed matrix of .

[0125] M2, intelligent prediction module, sequentially performs time alignment, normalization, and splicing operations on the standardized data stream and the external environment data stream to generate a splicing vector; divides the splicing vector into a second sample to generate a power prediction data set; uses the power prediction data set to train a second deep learning model, and when the second deep learning model converges, it is used to predict the power demand and power cost in a fixed time period in the future; Figure 2 .

[0126] In this embodiment, the time alignment is to unify the timestamps of the standardized data stream and the external environment data stream; the normalization is to normalize the values ​​of the standardized data stream and the external environment data stream to the interval [0,1] using the Z-score algorithm.

[0127] In this embodiment, the formula for the splicing operation is as follows:

[0128]

[0129] Among them, Y t represents the splicing vector, S t is the external environment data stream; Concat() is the concatenation function. In this embodiment, the external environment data stream is represented as:

[0130] S t =[S t, 1,S t, 2,...,S t,s ]

[0131] Among them S t,s is the sth external environment data at time t; s is the number of external environment data.

[0132] In some embodiments, the external environmental variables are electricity price, electricity demand, and the like.

[0133] In this embodiment, the length of the second sample is q, and the second sample represents [Y t-q ,Y t-q+1 ,...,Y t-1 ].

[0134] In this embodiment, the second deep learning model takes the second sample as input and outputs the power demand and power cost in a fixed time period;

[0135] The training process of the second deep learning model is as follows:

[0136] The first layer is the input layer: receiving the second sample;

[0137] The second layer is an embedding layer: the second sample is linearly transformed into an embedding vector, and the length of the embedding vector is equal to the length of the second sample.

[0138] The third layer is a Bi-GRU layer: it is used to extract the bidirectional time series features of the embedding vector to generate a bidirectional time series feature vector.

[0139] The fourth layer is the output layer: it contains the power demand prediction branch and the power cost prediction branch, which are used to output the power demand and power cost in a fixed time period in the future respectively.

[0140] The power demand forecasting branch adopts linear transformation forecasting, and the power cost forecasting branch adopts nonlinear transformation. The formula is as follows:

[0141]

[0142] in, represents the power demand in the future H time period, W represents the electricity cost in the future H time period. D 、W C1 and W C2 is the coefficient weight, b D 、b C1 and b C2 is the bias weight.

[0143] The loss function of the second deep learning prediction model includes a demand loss term and a cost loss term. The formula for the demand loss term is as follows:

[0144]

[0145] in, represents the predicted power demand at time t+h of the g-th second sample, represents the actual power demand at time t+h of the g-th second sample; represents the predicted power demand at time t+h of the g-th second sample, represents the actual power demand at time t+h of the g-th second sample; G represents the number of the second samples.

[0146] The loss function of the second deep learning model is expressed as: L1=L D +λL C , λ is the balance coefficient.

[0147] After the training of the second deep learning model is completed, the real standardized data stream and the external environment stream are input into the second deep learning model to predict the power demand and power cost in the future time period.

[0148] In some embodiments, H=24 hours, q=168 hours, that is, the standardized data stream of the past 7 days and the external environment data stream are used to predict the power demand and power cost for the next day.

[0149] M3, strategy optimization module, receives the predicted power demand and power cost and the external environment data stream, and generates energy control strategy in combination with reinforcement learning algorithm. Figure 3 .

[0150] In some embodiments, the reinforcement learning training process is as follows:

[0151] Define the state space as:

[0152]

[0153] Among them, S t,1 represents the electricity price at time t;

[0154] Define the action space as:

[0155] a t ={U t ,E t}

[0156] Among them, U t represents the power equipment dispatch vector at time t, E t Represents the energy storage charging and discharging power of the power equipment at time t.

[0157] The reward function is defined as:

[0158]

[0159] In this embodiment, the state z at time t-1 is t-1 Input to the Actor decision network to generate action a at time t t , and collect the state z at time t t , change action a t and state z t Input into the Critic value network to judge action a t The value of , the Critic value network contains the reward function.

[0160] The reward function includes an electricity cost reward item, an electricity equipment scheduling reward item and an electricity equipment energy storage charging and discharging reward item. The purpose of the electricity cost reward item is to minimize the difference between the electricity cost and the actual electricity cost. The purpose of the electricity equipment scheduling reward item is to minimize the number of electricity equipment dispatches. The purpose of the electricity equipment energy storage charging and discharging power reward item is to minimize the change in energy storage charging and discharging power.

[0161] The Critic value network is optimized by the mean square error function until convergence to maximize the reward; the Actor decision network is optimized by the policy gradient algorithm until convergence. After the Critic value network and the Actor decision network converge, the optimal reinforcement learning algorithm is obtained;

[0162] The optimal reinforcement learning algorithm receives the predicted power demand and power cost and the external environment data stream to generate an energy control strategy, which is equivalent to the action of the reinforcement algorithm.

[0163] A big data-driven intelligent energy management method, the method comprising the following steps:

[0164] Analyzing the power equipment parameters using the trained first deep learning model to generate correlation weights; and combining the correlation weights and the power equipment parameters to generate a standardized data stream;

[0165] The standardized data stream and the external environment data stream are sequentially time-aligned, normalized, and spliced ​​to generate a spliced ​​vector; the spliced ​​vector is divided into a second sample to generate a power prediction data set; the power prediction data set is used to train a second deep learning model, and when the second deep learning model converges, the model is used to predict power demand and power cost;

[0166] The predicted power demand and power cost and the external environment data stream are received, and an energy control strategy is generated in combination with a reinforcement learning algorithm.

[0167] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A big data driven intelligent energy management system, characterized in that: include: A feature extraction module that uses the trained first deep learning model to analyze power equipment parameters to generate correlation weights; Combine the associated weights and power equipment parameters to generate a standardized data stream; An intelligent prediction module sequentially performs time alignment, normalization, and splicing operations on the standardized data stream and the external environment data stream to generate a splicing vector; dividing the concatenated vector into second samples to generate a power prediction data set; Using the power prediction dataset to train a second deep learning model, and after the second deep learning model has converged, using it to predict power demand and power cost; A strategy optimization module receives the predicted power demand and power cost and the external environment data stream, and generates an energy control strategy in combination with a reinforcement learning algorithm.

2. The system according to claim 1, wherein: The power equipment parameters are expressed in the form of parameter time series, which is as follows: X t =[X t,1 ,X t,2 ,...,X t,I ] where X t,i is the parameter of the i-th power equipment at time t; I is the number of power equipment parameters; The parameter time series of the historical power equipment parameters is cut into N first samples to generate a power equipment parameter data set, each first sample has a length of p, and the first sample is represented by [X t-p ,X t-p+1 ,...,X t-1 ]; The training process of the first deep learning model is as follows: The first layer is the input layer: it receives the first sample; The second layer is the time series analysis layer: the first sample is input into the LSTM unit, and the time series feature vector is obtained by combining the global average pooling operation; The third layer is the parameter association analysis layer: the first sample is simultaneously input into the complete regression model and the residual regression model to predict the power equipment parameters at the next moment; Define parameter relationship matrix M={M ij }, matrix element M ij Represents the relationship coefficient between the i-th power equipment parameter and the j-th power equipment parameter; For the matrix element M ij The formula of the established complete regression model is as follows: in, In the complete regression model, the jth power equipment parameter at time t predicted by the ith power equipment parameter is X t,-i represents the parameters of other power equipment except the parameters of the i-th power equipment at time t; α j is the intercept term; β jk and γ jk is the regression coefficient;∈ t is the random error term; For the matrix element M ij The formula of the residual regression model established is as follows: in, In the residual regression model, it represents the jth power equipment parameter at time t predicted by the i-th power equipment parameter; Calculate the F statistic using the formula: When F>F α (p, N-2p-1), it is considered that the parameter i of the i-th power equipment has a Granger causal relationship with the parameter j of the j-th power equipment, and the M ij =β jk , Among them, F α (p,N-2p-1) represents the quantile when the cumulative probability on the right is α in the F distribution with (p,N-2p-1) degrees of freedom; in, The fourth layer is the output layer: the attention mechanism is used to calculate the temporal feature score and parameter relationship score respectively. The formulas are as follows: a trend =Softmax(W trend T+b trend ) a rel =Softmax(W rel M+b rel ) Among them, W trend 、W rel is the weight coefficient, b trend and b rel is the bias term; The temporal feature score and the parameter relationship score are combined to obtain the association weight Ω, which is as follows: Ω=a trend ⊙T+a rel ⊙M Where, ⊙term-by-term product operator = {Ω ij }, where Ω ij represents the association weight between power equipment parameter i and power equipment parameter j; The dimension of the time feature vector is the same as the dimension of the parameter association matrix, so the number of LSTM units is equal to the square of the number of power equipment parameters.

3. The system according to claim 2, characterized in that The loss function of the first deep learning model is: in, is the weighted prediction value, After the first deep learning model training converges, the optimal association weight is used; the optimal association weight is multiplied by the transposed matrix of the power equipment parameter to generate a standardized data stream The formula is as follows: in, Represents X t The transposed matrix of .

4. The system according to claim 1 or 3, characterized in that The formula for the splicing operation is as follows: Among them, Y t represents the splicing vector, S t For external environment data flow; The external environment data flow is represented as: S t =[S t,1 ,S t, 2,...,S t,s ] Among them S t,s is the sth external environment data at time t; s is the number of external environment data.

5. The system according to claim 4, characterized in that The second sample length is q, and the second sample represents [Y t-q ,Y t-q+1 ,...,Y t-1 ]; The second deep learning model takes the second sample as input, and the training process is as follows: The first layer is the input layer: receiving the second sample; The second layer is an embedding layer: the second sample is linearly transformed into an embedding vector, and the length of the embedding vector is equal to the length of the second sample; The third layer is a Bi-GRU layer: used to extract the bidirectional time series features of the embedded vector to generate a bidirectional time series feature vector; The fourth layer is the output layer: it contains the power demand forecast branch and the power cost forecast branch, which are used to output the power demand and power cost of a fixed time period in the future respectively; The power demand forecasting branch adopts linear transformation forecasting, and the power cost forecasting branch adopts nonlinear transformation forecasting, and the formulas are respectively expressed as follows: in, represents the power demand in the future H time period, W represents the electricity cost in the future H time period; D 、W C1 and W C2 is the coefficient weight, b D 、b C1 and b C2 is the bias weight.

6. The system according to claim 5, characterized in that The loss function of the second deep learning prediction model includes a demand loss term and a cost loss term. The formula for the demand loss term is as follows: in, represents the predicted power demand at time t+h of the g-th second sample, represents the actual power demand at time t+h of the g-th second sample; represents the predicted power demand at time t+h of the g-th second sample, represents the actual power demand at time t+h of the g-th second sample; G represents the number of the second samples; The loss function of the second deep learning model is expressed as: L1=L D +λL C , λ is the balance coefficient.

7. The system according to claim 6, characterized in that The training process of reinforcement learning is as follows: Define the state space as: Among them, S t,1 The sth external environmental data at time t is the electricity price; The action space of the energy control strategy is defined as: the t ={U t ,AND t } Among them, U t represents the power equipment dispatch vector at time t, E t Indicates the energy storage charging and discharging power of the power equipment at time t; Define a reward function, which is to minimize the difference between the electricity cost and the actual electricity cost, minimize the number of power equipment dispatches, and minimize the number of power equipment dispatches. There are power cost reward items, power equipment dispatch reward items, and power equipment energy storage charging and discharging reward items, respectively. The formula is as follows: The state z at time t-1 t-1 Input to the Actor decision network to generate action a at time t t , and collect the state z at time t t , change action a t and state z t Input into the Critic value network to judge action a t The value of , the Critic value network includes the reward function; Optimize the Critic value network through the mean square error function until convergence to maximize the reward; optimize the Actor decision network through the policy gradient algorithm until convergence; After the Critic value network and the Actor decision network converge, the optimal reinforcement learning algorithm is obtained.

8. The system according to claim 7, characterized in that The optimal reinforcement learning algorithm receives the predicted power demand and power cost and the external environment data stream to generate actions of an energy control strategy.

9. A big data driven intelligent energy management method, characterized in that: The following steps are involved: Analyzing power equipment parameters using the trained first deep learning model to generate associated weights; Combine the associated weights and power equipment parameters to generate a standardized data stream; The standardized data stream and the external environment data stream are sequentially time-aligned, normalized, and spliced ​​to generate a splicing vector; dividing the concatenated vector into second samples to generate a power prediction data set; Using the power prediction dataset to train a second deep learning model, and after the second deep learning model has converged, using it to predict power demand and power cost; The predicted power demand and power cost and the external environment data stream are received, and an energy control strategy is generated in combination with a reinforcement learning algorithm.

Images