A big data driven intelligent energy management system and method

Abstract

The application discloses a kind of big data driven intelligent energy management system and method, belong to energy management technical field.System includes: feature extraction module, the Granger causality between power equipment parameters is analyzed by first deep learning model and time sequence feature, generates associated weight matrix and constructs standardization data stream;Intelligent prediction module, standardization data stream is fused with external environment data, after time alignment, normalization input second deep learning model, synchronously predict power demand and cost;Strategy optimization module, based on the prediction result design reinforcement learning framework, to minimize cost, equipment scheduling times and charge-discharge fluctuation as target generate control strategy.The application improves the feature interpretability through causal analysis, bidirectional time series network enhances prediction accuracy, and adopts reinforcement learning to realize dynamic optimization closed loop, significantly improves the intelligent level of energy management and resource utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of big data, specifically to a big data-driven intelligent energy management system and method. Background Technology

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

[0003] The core bottlenecks of existing technologies are: ① The mining of correlation features of power equipment parameters lacks an interpretable framework, making it difficult to reveal physical coupling relationships through causal tests; ② The spatiotemporal alignment and feature fusion mechanisms of multi-source data are imperfect, resulting in limited generalization ability of prediction models; ③ Energy control strategies lack dynamic optimization closed loops, making it impossible to adaptively adjust equipment operating states based on real-time prediction results. These problems directly lead to insufficient prediction accuracy and lagging control strategies in energy management systems, restricting the improvement of energy utilization efficiency and the optimization of overall costs. Therefore, it is urgent to construct an intelligent energy management system that integrates big data analysis, deep learning, and reinforcement learning to achieve intelligent processing from data feature analysis to strategy generation. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes a big data-driven intelligent energy management system and method.

[0005] Therefore, the technical solution adopted by the present invention is as follows:

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

[0007] The feature extraction module uses the first deep learning model after training to analyze the parameters of the power equipment to generate correlation weights; and combines the correlation weights and the parameters of the power equipment to generate a standardized data stream.

[0008] The intelligent prediction module performs time alignment, normalization, and concatenation operations on the standardized data stream and the external environment data stream in sequence to generate a concatenated vector; the concatenated vector is then segmented into a second sample to generate a power prediction dataset; the power prediction dataset is used to train a second deep learning model, and once the second deep learning model has converged, it is used to predict power demand and power costs.

[0009] The strategy optimization module receives the predicted electricity demand and electricity cost, as well as the external environment data stream, and generates an energy control strategy by combining it with a reinforcement learning algorithm.

[0010] Furthermore, the power equipment parameters are represented in the form of a parameter time series, as follows:

[0011]

[0012] in Let I be the parameter of the i-th power equipment at time t; I is the number of power equipment parameters.

[0013] The historical power equipment parameter time series is divided into N first samples to generate a power equipment parameter dataset. Each first sample has a length of p and is represented as follows: ;

[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 T is obtained by combining the global average pooling operation;

[0017] The third layer is the parameter correlation 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 the parameter relationship matrix , matrix elements This represents the relationship coefficient between the i-th power equipment parameter and the j-th power equipment parameter;

[0019] For matrix elements The formula for the established complete regression model is as follows:

[0020]

[0021] in, This represents the predicted parameter of the j-th power equipment at time t in the complete regression model, where the parameter of the i-th power equipment is the same as the parameter of the j-th power equipment at time t. This represents the parameters of the power equipment other than the i-th power equipment parameter at time tk; For the intercept term; and These are the regression coefficients; This is the random error term;

[0022] For matrix elements The formula for the established residual regression model is as follows:

[0023]

[0024] in, This represents the predicted parameter of the j-th power equipment at time t in the residual regression model, where the parameter of the i-th power equipment is the parameter of the j-th power equipment at time t.

[0025] The F-statistic is calculated using the following formula:

[0026]

[0027] when At that time, it is assumed that there is a Granger causality relationship between parameter i of the i-th power equipment and parameter j of the j-th power equipment, and the value is updated. ,

[0028] in, This represents the quantile on the right side of an F-distribution with cumulative probability α, where the distribution follows a distribution with (p, N−2p−1) degrees of freedom.

[0029] in, ; ;

[0030] The fourth layer is the output layer: it uses an attention mechanism to calculate the temporal feature score and the parameter relationship score, respectively, with the following formulas:

[0031]

[0032]

[0033] in, , These are the weighting coefficients. and For bias terms;

[0034] The association weight is obtained by fusing temporal feature scores and parameter relationship scores. The formula is as follows:

[0035]

[0036] in, This is the term-by-term multiplication operator;

[0037] The dimension of the time feature vector is the same as the dimension of the parameter correlation matrix, therefore 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, This is a weighted forecast value. ;

[0041] After the first deep learning model converges during training, it is used to obtain the optimal correlation weights; the optimal correlation weights are then multiplied with the transpose of the power equipment parameters to generate a standardized data stream. The formula is as follows:

[0042]

[0043] in, express The transpose of .

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

[0045]

[0046] in, This represents the concatenated vector. For external environment data flow;

[0047] The external environment data stream is represented as follows:

[0048]

[0049] in Let be the s-th external environment data at time t; s is the number of external environment data.

[0050] Furthermore, the second sample has a length of q, and the second sample represents... ;

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

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

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

[0054] The third layer is the Bi-GRU layer: used to extract the bidirectional temporal features of the embedded vector to generate a bidirectional temporal feature vector.

[0055] The fourth layer is the output layer: it contains an electricity demand forecasting branch and an electricity cost forecasting branch, which are used to output the electricity demand and electricity cost for a fixed future time period, respectively.

[0056] The electricity demand forecasting branch uses linear transformation forecasting, while the electricity cost forecasting branch uses nonlinear transformation forecasting. The formulas are expressed as follows:

[0057]

[0058]

[0059] in, This represents the electricity demand over the next H time period. This represents the electricity cost over the next H time period; , and For coefficient weights, , and This is the bias weight.

[0060] Furthermore, the loss function of the second deep learning prediction model includes a demand loss term and a cost loss term, the formula for which is as follows:

[0061]

[0062]

[0063] in, Let represent the predicted electricity demand at time t+h for the g-th second sample. Let represent the actual electricity demand at time t+h for the g-th second sample; This represents the predicted electricity cost at time t+h for the g-th second sample. Let G represent the actual electricity cost at time t+h for the g-th second sample; G represents the number of second samples.

[0064] The loss function of the second deep learning model is expressed as: , This is the balance coefficient.

[0065] Furthermore, the training process of the reinforcement learning is as follows:

[0066] Define the state space as follows:

[0067]

[0068] in, Let represent the electricity price at time t; This represents the electricity demand over the next H time period. This represents the electricity cost over the next H time period;

[0069] Define the action space of the energy control strategy as follows:

[0070]

[0071] in, This represents the power equipment scheduling vector at time t. This represents the energy storage charging and discharging power of the power equipment at time t;

[0072] Define a reward function that minimizes the difference between the actual electricity cost and the power cost, minimizes the number of power equipment dispatches, and minimizes the number of power equipment dispatches. The reward function includes an electricity cost reward, a power equipment dispatch reward, and a power equipment energy storage charging / discharging reward, as shown in the following formula:

[0073]

[0074]

[0075] The state at time t-1 The input is fed into the Actor decision network to generate the action at time t. And collect the state at time t. , will the action and state Input into the Critic value network to evaluate actions The value of the Critic value network includes the reward function;

[0076] The Critic value network is optimized using the mean squared error function until convergence to maximize rewards; the Actor decision network is optimized using the policy gradient algorithm until convergence.

[0077] After the Critic value network and Actor decision network converge, the optimal reinforcement learning algorithm is obtained.

[0078] Furthermore, the optimal reinforcement learning algorithm receives the predicted electricity demand and electricity cost, as well as the external environment data stream, to generate actions for an energy control strategy.

[0079] This invention also proposes a big data-driven intelligent energy management method, characterized by the following steps:

[0080] The first deep learning model after training is used to analyze the parameters of the power equipment to generate correlation weights; the correlation weights and the parameters of the power equipment are combined to generate a standardized data stream.

[0081] The standardized data stream and the external environment data stream are sequentially time-aligned, normalized, and concatenated to generate a concatenated vector; the concatenated vector is then segmented into a second sample to generate a power prediction dataset; the power prediction dataset is used to train a second deep learning model, and once the second deep learning model has converged, it is used to predict power demand and power costs.

[0082] The system receives the predicted electricity demand and electricity cost, as well as the external environment data stream, and generates an energy control strategy by combining it with a reinforcement learning algorithm.

[0083] Compared with the prior art, the advantages of the present invention are as follows:

[0084] This invention analyzes the physical correlation between power equipment parameters through causal testing and attention mechanisms, and captures operating trends and periodic patterns by combining time series analysis networks, generating an interpretable dynamic weight matrix to achieve in-depth mining of parameter correlation features.

[0085] This invention integrates power equipment parameters and external environmental data to construct a multi-dimensional input feature space, and extracts bidirectional dependencies from historical data through a bidirectional time-series network. It also employs two task prediction branches to simultaneously output power demand and cost, providing comprehensive and reliable forward-looking information for energy planning.

[0086] This invention utilizes a reinforcement learning model to dynamically optimize equipment scheduling and energy storage strategies to balance costs. Combined with a real-time feedback closed-loop optimization mechanism, the system can automatically adapt to fluctuations in energy supply and demand, achieving intelligent management throughout the entire process from prediction to control. Attached Figure Description

[0087] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0088] Figure 1 This is a schematic diagram of the feature extraction module of the present invention.

[0089] Figure 2 This is a schematic diagram of the intelligent prediction module of the present invention.

[0090] Figure 3 This is a schematic diagram of the strategy optimization module of the present invention. Detailed Implementation

[0091] To achieve the above objectives, the present invention provides a big data-driven intelligent energy management system, comprising:

[0092] M1, the feature extraction module, uses a first deep learning model to analyze power equipment parameters to generate correlation weights; combining the correlation weights and power equipment parameters, it generates a standardized data stream; see... Figure 1 .

[0093] In this embodiment, the power equipment parameters are represented in the form of a parameter time series, which is expressed as follows: ,in Let be the parameter of the i-th power equipment at time t; I is the number of power equipment parameters.

[0094] In this embodiment, the time series of historical power equipment parameters is divided into N first samples to generate a power equipment parameter dataset. Each first sample has a length of p, and the first sample is represented as... The first deep learning model is trained based on the power equipment parameter dataset.

[0095] In this embodiment, the first deep learning model takes the first sample as input and the association weight as output;

[0096] The training process of the first deep learning model is as follows:

[0097] The first layer is the input layer: it receives the first sample.

[0098] The second layer is the time series analysis layer: the first sample is input into the LSTM unit to obtain the time series feature vector; the LSTM unit consists of an input gate, a forget gate, and an output gate, and the final output state is the hidden state vector. The detailed structure of LSTM will not be described in detail here.

[0099] The average value of the hidden state features at all time points is calculated using global average pooling to obtain the temporal feature vector T.

[0100] The third layer is the parameter correlation 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 time step; at the same time, a parameter relationship matrix is ​​defined. , matrix elements This represents the relationship coefficient between the i-th power equipment parameter and the j-th power equipment parameter.

[0101] For matrix elements The formula for the established complete regression model is as follows:

[0102]

[0103] in, This represents the predicted parameter of the j-th power equipment at time t in the complete regression model, where the parameter of the i-th power equipment is the same as the parameter of the j-th power equipment at time t. This represents the parameters of the power equipment other than the i-th power equipment parameter at time tk; For the intercept term; and These are the regression coefficients; This is the random error term.

[0104] For matrix elements The formula for the established residual regression model is as follows:

[0105]

[0106] in, This represents the predicted parameter of the j-th power equipment at time t in the residual regression model, where the parameter of the i-th power equipment is the parameter of the j-th power equipment at time t.

[0107] The F-statistic is calculated using the following formula:

[0108]

[0109] when At that time, it is assumed that there is a Granger causality relationship between parameter i of the i-th power equipment and parameter j of the j-th power equipment, and the value is updated. ,

[0110] in, Let denot α, which represents the quantile on the right side of an F-distribution with degrees of freedom (p, N−2p−1) and a cumulative probability of α.

[0111] in, ; .

[0112] The fourth layer is the output layer: it uses an attention mechanism to calculate the temporal feature score and the parameter relationship score, respectively, with the following formulas:

[0113]

[0114]

[0115] in, For the Softmax function, , These are the weighting coefficients. and This is a bias term.

[0116] The association weight parameters are obtained by fusing time feature scores and parameter relationship scores. The formula is as follows:

[0117]

[0118] in, This is the term-by-term multiplication operator.

[0119] In this embodiment, the dimension of the time feature vector is the same as the dimension of the parameter correlation matrix. Therefore, it is necessary to ensure that the number of LSTM units is equal to the square of the number of power equipment parameters, i.e., I2.

[0120] The loss function of the first deep learning model is:

[0121]

[0122] in, This is a weighted forecast value. .

[0123] The loss function is used to optimize the coupling relationship between power equipment parameters.

[0124] In some embodiments, electrical equipment parameters include current, voltage, and temperature.

[0125] After the first deep learning model converges, the optimal correlation weights are obtained. The optimal correlation weights are then multiplied with the transpose of the power equipment parameters to generate a standardized data stream. The formula is as follows:

[0126]

[0127] in, express The transpose of .

[0128] M2, the intelligent prediction module, sequentially performs time alignment, normalization, and concatenation operations on the standardized data stream and the external environment data stream to generate a concatenated vector; the concatenated vector is then segmented into a second sample to generate a power prediction dataset; this power prediction dataset is used to train a second deep learning model, and once the second deep learning model converges, it is used to predict power demand and power costs for a fixed future time period; see [link / reference]. Figure 2 .

[0129] 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 use the Z-score algorithm to normalize the values ​​of the standardized data stream and the external environment data stream to the [0,1] interval.

[0130] In this embodiment, the formula for the splicing operation is as follows:

[0131]

[0132] in, This represents the concatenated vector. This refers to the external environment data stream; Concat() is the concatenation function.

[0133] In this embodiment, the external environment data stream is represented as:

[0134]

[0135] in Let be the s-th external environment data at time t; s is the number of external environment data.

[0136] In some embodiments, external environmental variables include electricity prices, electricity demand, etc.

[0137] In this embodiment, the second sample length is q, and the second sample represents... .

[0138] In this embodiment, the second deep learning model takes the second sample as input and the electricity demand and electricity cost over a fixed time period as output.

[0139] The training process of the second deep learning model is as follows:

[0140] The first layer is the input layer: it receives the second sample;

[0141] The second layer is the embedding layer: the second sample is linearly transformed into an embedding vector, the length of which is equal to the length of the second sample.

[0142] The third layer is the Bi-GRU layer: used to extract the bidirectional temporal features of the embedded vector to generate a bidirectional temporal feature vector.

[0143] The fourth layer is the output layer: it contains an electricity demand forecasting branch and an electricity cost forecasting branch, which are used to output the electricity demand and electricity cost for a fixed future time period, respectively.

[0144] The electricity demand forecasting branch uses linear transformation, while the electricity cost forecasting branch uses nonlinear transformation, as expressed in the following formulas:

[0145]

[0146]

[0147] in, This represents the electricity demand over the next H time period. This represents the electricity cost over the next H time period. , and For coefficient weights, , and This is the bias weight.

[0148] The loss function of the second deep learning prediction model includes a demand loss term and a cost loss term, and the formula for the demand loss term is as follows:

[0149]

[0150]

[0151] in, Let represent the predicted electricity demand at time t+h for the g-th second sample. Let represent the actual electricity demand at time t+h for the g-th second sample; This represents the predicted electricity cost at time t+h for the g-th second sample. Let G represent the actual electricity cost at time t+h for the g-th second sample; G represents the number of the second samples.

[0152] The loss function of the second deep learning model is expressed as: , This is the balance coefficient.

[0153] Once the second deep learning model is trained, the real standardized data stream and external environment stream are input into the second deep learning model to predict the electricity demand and electricity cost for future periods.

[0154] In some embodiments, H = 24 hours and q = 168 hours, that is, using standardized data streams from the past 7 days and external environment data streams to predict electricity demand and electricity costs for the next day.

[0155] M3, the strategy optimization module, receives the predicted electricity demand and cost, as well as the external environment data stream, and generates an energy control strategy using a reinforcement learning algorithm. (See...) Figure 3 .

[0156] In some embodiments, the training process of the reinforcement learning is as follows:

[0157] Define the state space as follows:

[0158]

[0159] in, Let represent the electricity price at time t;

[0160] Define the action space as follows:

[0161]

[0162] in, This represents the power equipment scheduling vector at time t. This represents the energy storage charging and discharging power of the power equipment at time t.

[0163] Define the reward function as follows:

[0164]

[0165]

[0166] In this embodiment, the state at time t-1 is... The input is fed into the Actor decision network to generate the action at time t. And collect the state at time t. , will the action and state Input into the Critic value network to evaluate actions The value of the Critic value network includes the reward function.

[0167] The reward function includes an electricity cost reward, an electricity equipment dispatch reward, and an electricity equipment energy storage charging and discharging reward. The purpose of the electricity cost reward is to minimize the difference between the electricity cost and the actual electricity cost. The purpose of the electricity equipment dispatch reward is to minimize the number of times the electricity equipment is dispatched. The purpose of the electricity equipment energy storage charging and discharging power reward is to minimize the change in energy storage charging and discharging power.

[0168] The Critic value network is optimized using the mean squared error function until convergence to maximize reward; the Actor decision network is optimized using the policy gradient algorithm until convergence. Once the Critic value network and Actor decision network converge, the optimal reinforcement learning algorithm is obtained.

[0169] The optimal reinforcement learning algorithm receives the predicted electricity demand and electricity cost, as well as the external environment data stream, to generate an energy control strategy, which is equivalent to the action of the reinforcement algorithm.

[0170] A big data-driven intelligent energy management method, comprising the following steps:

[0171] The first deep learning model after training is used to analyze the parameters of the power equipment to generate correlation weights; the correlation weights and the parameters of the power equipment are combined to generate a standardized data stream.

[0172] The standardized data stream and the external environment data stream are sequentially time-aligned, normalized, and concatenated to generate a concatenated vector; the concatenated vector is then segmented into a second sample to generate a power prediction dataset; the power prediction dataset is used to train a second deep learning model, and once the second deep learning model has converged, it is used to predict power demand and power costs.

[0173] The system receives the predicted electricity demand and electricity cost, as well as the external environment data stream, and generates an energy control strategy by combining it with a reinforcement learning algorithm.

[0174] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A big data-driven intelligent energy management system, characterized in that, include: The feature extraction module uses the first deep learning model after training to analyze the parameters of the power equipment in order to generate correlation weights; By combining correlation weights and power equipment parameters, a standardized data stream is generated; The intelligent prediction module performs time alignment, normalization, and splicing operations on the standardized data stream and the external environment data stream in sequence to generate a spliced ​​vector. The concatenated vector is split into a second sample to generate a power prediction dataset; The second deep learning model is trained using the power forecasting dataset. Once the second deep learning model has converged, it is used to predict power demand and power costs. The strategy optimization module receives the predicted power demand and power cost, as well as the external environment data stream, and generates an energy control strategy by combining it with a reinforcement learning algorithm. The power equipment parameters are represented in time series form, as follows: in Let be the parameters of the i-th power equipment at time t; I represents the number of parameters for the power equipment; The historical power equipment parameter time series is divided into N first samples to generate a power equipment parameter dataset. Each first sample has a length of p and is represented as follows: ; 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 T is obtained by combining the global average pooling operation; The third layer is the parameter correlation 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 the parameter relationship matrix , matrix elements This represents the relationship coefficient between the i-th power equipment parameter and the j-th power equipment parameter; For matrix elements The formula for the established complete regression model is as follows: in, This represents the predicted parameter of the j-th power equipment at time t in the complete regression model, where the parameter of the i-th power equipment is the same as the parameter of the j-th power equipment at time t. This represents the parameters of the power equipment other than the i-th power equipment parameter at time tk; For the intercept term; and These are the regression coefficients; This is the random error term; For matrix elements The formula for the established residual regression model is as follows: in, This represents the predicted parameter of the j-th power equipment at time t in the residual regression model, where the parameter of the i-th power equipment is the parameter of the j-th power equipment at time t. The F-statistic is calculated using the following formula: when At that time, it is assumed that there is a Granger causality relationship between parameter i of the i-th power equipment and parameter j of the j-th power equipment, and the value is updated. , in, This represents the quantile on the right side of an F-distribution with cumulative probability α, where the distribution follows a distribution with degrees of freedom of (p, N−2p−1). in, ; ; The fourth layer is the output layer: it uses an attention mechanism to calculate the temporal feature score and the parameter relationship score, respectively, with the following formulas: in, , These are the weighting coefficients. and For bias terms; The association weight is obtained by fusing temporal feature scores and parameter relationship scores. The formula is as follows: in, This is the term-by-term multiplication operator; The dimension of the time feature vector is the same as the dimension of the parameter correlation matrix, therefore the number of LSTM units is equal to the square of the number of power equipment parameters.

2. The system according to claim 1, characterized in that, The loss function of the first deep learning model is: in, This is a weighted forecast value. ; After the first deep learning model converges during training, it is used to obtain the optimal correlation weights; the optimal correlation weights are then multiplied with the transpose of the power equipment parameters to generate a standardized data stream. The formula is as follows: in, express The transpose of .

3. The system according to claim 1 or 2, characterized in that, The formula for the splicing operation is as follows: in, This represents the concatenated vector. For external environment data flow; The external environment data stream is represented as follows: in Let be the s-th external environment data at time t; s is the number of external environment data.

4. The system according to claim 3, characterized in that, The second sample has a length of q, and the second sample represents ; The second deep learning model uses the second sample as input, and the training process is as follows: The first layer is the input layer: it receives the second sample; The second layer is the embedding layer: the second sample is linearly transformed into an embedding vector, the length of which is equal to the length of the second sample; The third layer is the Bi-GRU layer: used to extract the bidirectional temporal features of the embedded vector to generate a bidirectional temporal feature vector; The fourth layer is the output layer: it contains an electricity demand forecasting branch and an electricity cost forecasting branch, which are used to output the electricity demand and electricity cost for a fixed future time period, respectively. The electricity demand forecasting branch uses linear transformation forecasting, while the electricity cost forecasting branch uses nonlinear transformation forecasting. The formulas are expressed as follows: in, This represents the electricity demand over the next H time period. This represents the electricity cost over the next H time period; , and For coefficient weights, , and This is the bias weight.

5. The system according to claim 4, characterized in that, The loss function of the second deep learning prediction model includes a demand loss term and a cost loss term, and the formula for the demand loss term is as follows: in, Let represent the predicted electricity demand at time t+h for the g-th second sample. Let represent the actual electricity demand at time t+h for the g-th second sample; This represents the predicted electricity cost at time t+h for the g-th second sample. Let G represent the actual electricity cost at time t+h for the g-th second sample; G represents the number of second samples. The loss function of the second deep learning model is expressed as: , This is the balance coefficient.

6. The system according to claim 5, characterized in that, The training process for reinforcement learning is as follows: Define the state space as follows: in, The s-th external environmental data at time t represents the electricity price; Define the action space of the energy control strategy as follows: in, This represents the power equipment scheduling vector at time t. This represents the energy storage charging and discharging power of the power equipment at time t; Define a reward function that minimizes the difference between the actual electricity cost and the power cost, minimizes the number of power equipment dispatches, and minimizes the number of power equipment dispatches. The reward function includes an electricity cost reward, a power equipment dispatch reward, and a power equipment energy storage charging / discharging reward, as shown in the following formula: The state at time t-1 The input is fed into the Actor decision network to generate the action at time t. And collect the state at time t. , will the action and state Input into the Critic value network to evaluate actions The value of the Critic value network includes the reward function; The Critic value network is optimized using the mean squared error function until convergence to maximize rewards; the Actor decision network is optimized using the policy gradient algorithm until convergence. After the Critic value network and Actor decision network converge, the optimal reinforcement learning algorithm is obtained.

7. The system according to claim 6, characterized in that, The optimal reinforcement learning algorithm receives the predicted electricity demand and electricity cost, as well as the external environment data stream, to generate actions for an energy control strategy.

8. A big data-driven intelligent energy management method, applied to the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: The parameters of the power equipment are analyzed using the first deep learning model after training to generate correlation weights; By combining correlation weights and power equipment parameters, a standardized data stream is generated; The standardized data stream and the external environment data stream are sequentially time-aligned, normalized, and concatenated to generate a concatenated vector; The concatenated vector is split into a second sample to generate a power prediction dataset; The second deep learning model is trained using the power forecasting dataset. Once the second deep learning model has converged, it is used to predict power demand and power costs. The system receives the predicted electricity demand and electricity cost, as well as the external environment data stream, and generates an energy control strategy by combining it with a reinforcement learning algorithm.

Images