Multi-scale multi-group cognitive intelligent scheduling and control method for integrated energy system

Abstract

The application provides a multi-scale multi-group cognitive intelligent scheduling and control method for a comprehensive energy system, which comprises a multi-time scale fusion control framework and two modules, the first module is a control module based on adaptive noise complete ensemble empirical mode decomposition and time series prediction, and the second module is a scheduling module based on adaptive noise complete ensemble empirical mode decomposition and bidirectional long short-term memory network prediction and fusing an artificial behavior cognitive knowledge matrix. The method takes a frequency deviation sequence of the comprehensive system, a regional control deviation sequence, user behavior and climate factors as inputs, and outputs an automatic generation control unit control instruction updated every 4 seconds and a non-automatic generation control unit scheduling instruction updated every 5 minutes and 60 minutes through the control module and the scheduling module respectively. The method can solve the problem of incoordination between generation control and economic scheduling in the comprehensive energy system, improve the control precision and reduce the frequency deviation.

Description

Technical Field

[0001] This invention belongs to the field of power generation control in integrated energy systems, and relates to a power generation and dispatch coordination control method based on artificial intelligence technology, which is applicable to the power generation and dispatch control of integrated energy systems. Background Technology

[0002] Renewable energy has significant uncertainties. As the penetration rate of renewable energy in integrated energy systems continues to increase, the demand for flexibility and accuracy in the operation and control of integrated energy systems also increases.

[0003] Traditional power systems employ a combination of optimization methods, control methods, and further optimization methods to address the issues of generator unit combination, economic dispatch, automatic generation control, and generation control command allocation, thereby achieving a dynamic balance between the active power generated and consumed by the power system. However, in integrated energy systems dominated by renewable energy sources, thermal power units constitute a very small proportion, transforming the combination problem into a matter of economic dispatch, automatic generation control, and generation control command allocation. In integrated energy systems, the uncertainty and strong randomness of renewable energy and emerging loads necessitate a comprehensive energy storage system. This system can store excess active power and draw upon it when needed. Traditional multi-timescale control methods sometimes fail to address energy storage response issues. When the response time of an energy storage device is shorter than the interval between control commands, a new control command may arrive, but the energy storage system, due to its slow response, may still be executing the previous command. This can lead to severe back-regulation, causing power imbalance in the integrated energy system, disrupting frequency stability, and impacting power quality.

[0004] Therefore, a multi-scale, multi-group cognitive intelligent scheduling and control method for integrated energy systems is proposed. Through time-scale fusion and real-time control, the method simultaneously schedules automatic and non-automatic generation control units using reinforcement learning, avoiding reverse scheduling and unifying the optimization and control problems. Summary of the Invention

[0005] This invention proposes a multi-scale, multi-group cognitive intelligent scheduling and control method for integrated energy systems. Its key feature is that it can issue regular power generation and scheduling control commands to the integrated energy system using a multi-response time-scale fusion control framework. It combines adaptive noise complete set empirical mode decomposition, time series prediction, bidirectional long short-term memory network prediction, and various reinforcement learning methods to improve control accuracy and power quality, reduce frequency deviation and power generation costs, and dynamically maintain a balance between power generation control and economic dispatch. The steps in its application are as follows:

[0006] Step (1): In region A, obtain the frequency deviation sequence of the integrated energy system over a period of time. and region Integrated energy system regional control error sequence ,in ; obtain Real-time user behavior information ,temperature ,humidity ,wind direction Wind speed ,precipitation and air pressure ;

[0007] Step (2): Apply adaptive noise complete set empirical mode decomposition to the frequency deviation sequence. Process it, and set the number of iterations to 1. ,Will Positive and negative Gaussian white noise pairs are added to the original signal respectively. In total, for The first new signal, The new signal sequence after adding white noise. for:

[0008] (1)

[0009] In the formula, The standard deviation of the noise; For the first Gaussian white noise that follows a standard normal distribution;

[0010] For each Perform empirical mode decomposition, in the first... After the first empirical mode decomposition, the th... Modal components , decomposed The modal components are summed and averaged to obtain :

[0011] (2)

[0012] In the formula, This is the first intrinsic mode component;

[0013] Calculate the first residual component :

[0014] (3)

[0015] exist Add to The auxiliary noise signal after empirical mode decomposition of the positive and negative Gaussian white noise pairs:

[0016] (4)

[0017] In the formula, For the first The new signal after adding auxiliary noise; For the function used in empirical mode decomposition;

[0018] For each Perform empirical mode decomposition, in the first... After the first empirical mode decomposition, the th... Modal components , decomposed The modal components are summed and averaged to obtain :

[0019] (5)

[0020] In the formula, This is the second intrinsic mode component;

[0021] Calculate the second residual component :

[0022] (6)

[0023] Step (3): Repeat the calculation in step (2) until the residual components can no longer be subjected to empirical mode decomposition, and then proceed with... The number of intrinsic mode components obtained after the iteration is Then the original signal Decomposed into:

[0024] (7)

[0025] In the formula, For the first One intrinsic mode component; For the first One residual component;

[0026] Step (4): Put all and As input, a time series prediction method is used to process the input signal. The time series prediction neural network has two gating mechanisms in its recurrent unit: a reset gate and an update gate. In this prediction, middle value of time For input, reset the door. The output at time t is:

[0027] (8)

[0028] In the formula, To reset the door Output at any moment; It is the sigmoid function; To reset the state weights of the gate; To reset the input weights of the gate; To reset the door offset; Step size; This represents the hidden layer state of the loop unit at the previous time step;

[0029] Update Gate The output at time t is:

[0030] (9)

[0031] In the formula, To update the door Output at any moment; To update the state weights of the gate; and To update the input weights of the gate; To update the door offset;

[0032] Loop body unit The activation state of the hidden layer at any given time is:

[0033] (10)

[0034] In the formula, For loop body unit The activation state of the hidden layer at any given moment; For activation functions; For Hadamard product; These are the input weights for the hidden layer; The weights of the hidden layer; For the bias of the hidden layer;

[0035] Calculate a loop body unit Output at time:

[0036] (11)

[0037] In the formula, For loop body unit The output at each moment also serves as the hidden state for the next moment;

[0038] Step (5): Repeat step (4) to obtain a total of The predicted signal is denoted as the nth prediction signal. The predicted signals are ,Will Classified according to the magnitude of fluctuation, within the same time period Within, if the number of extreme points of the function graph Greater than or equal to the set value If the number of extreme points on the function graph is large, it is classified as a large fluctuation signal; Less than the set value If so, it is classified as a small fluctuation signal; for For large fluctuation signals, a "state-action-reward-state-action" learning method is used for reinforcement learning; for For small fluctuation signals, because the function changes relatively slowly, a fractional-order "proportional-integral-differential" method is used for learning and following; in the "state-action-reward-state-action" reinforcement learning method, the state... Next action Expected reward:

[0039] (12)

[0040] In the formula, For state Next action Expected rewards; The expected operation function; for Discount factor of time; for Momentary rewards; for The state at any given moment; for Actions at any given moment; To reinforce the learning state, i.e., the system frequency deviation ; This refers to the system's action, i.e., the unit's power output command;

[0041] After undergoing "state-action-reward-state-action" reinforcement learning, the large fluctuation signal outputs a processed signal for the large fluctuation signal. ;

[0042] In the fractional proportional-integral-differential method, the transfer function for:

[0043] (13)

[0044] In the formula, This is the proportional adjustment coefficient; This is the integral adjustment coefficient; This is the differential adjustment coefficient; For the order of integration; For the order of the differential and ;

[0045] The processing signal for small fluctuation signals is:

[0046] (14)

[0047] In the formula, This is a processed signal for small fluctuation signals; It is the inverse Laplace transform function; The Laplace transform of a small fluctuation signal;

[0048] Step (6): and The sum is used as the unit power output command, and the output of the automatic generation control unit is controlled in real time every 4 seconds.

[0049] Step (7): Combine the results obtained in step (1) with the region Integrated energy system regional control error sequence As input, the decomposed signal is obtained through adaptive noise complete set empirical analysis. The decomposed signal is predicted using a bidirectional long short-term memory network; the recurrent unit of the long short-term memory neural network has three gated units: a forget gate, an input gate, and an output gate, using the first decomposed signal as an example. middle value of time For input, forget gate The output at time t is:

[0050] (15)

[0051] In the formula, Forgotten Gate Output at any moment; The input weights for the forget gate; State weights for the forget gate; This refers to the external state and output of the previous time step; For the offset of the forget gate;

[0052] Input gate The output at time t is:

[0053] (16)

[0054] In the formula, For input gate Output at any moment; The input weights are the input weights of the input gate; The state weights are the input gate weights; For the input gate bias;

[0055] Candidate state The output at time t is:

[0056] (17)

[0057] In the formula, Candidate state Output at any moment; The input weights for the candidate states; The state weights of the candidate states; The bias for the candidate state;

[0058] Output gate The output at time t is:

[0059] (18)

[0060] In the formula, For output gate Output at any moment; The input weights are the values ​​for the output gate. The state weights of the output gate; For the output gate bias;

[0061] The internal state at time t is:

[0062] (19)

[0063] In the formula, for The internal state at any given moment; This refers to the internal state at the previous moment;

[0064] at last, The external state and output at time t are:

[0065] (20)

[0066] In the formula, for The external state and output at a given moment also serve as the input for the next moment;

[0067] Step (8): Repeat the process of step (7) to obtain the following: As the initial input prediction signal, the prediction signal is used to obtain the dispatch command by applying the consistency and other incremental rate criteria. The dispatch command is updated every 5 minutes for the non-automatic generation control unit.

[0068] The output allocation formula for the consistency-equal incremental rate criterion is as follows:

[0069] (twenty one)

[0070] In the formula, for Time crew Distributed load; for The incremental rate of power consumption in the time-of-flight system; and For the unit The coefficient of the consumption characteristic curve;

[0071] Step (9): exist value of time User behavior information ,temperature ,humidity ,wind direction Wind speed ,precipitation and air pressure As matrix elements, they form an information matrix. ;

[0072] (twenty two)

[0073] Construct a Artificial behavior cognitive knowledge matrix ,in The number of non-automatic generator control units is used, and the alternating direction multiplier method is employed to determine the knowledge matrix of human behavior cognition. Update each parameter in the data;

[0074] Information matrix With artificial behavior cognitive knowledge matrix Multiplication yields the output matrix :

[0075] (twenty three)

[0076] Output matrix Each element in As dispatch instructions, the corresponding non-automatic generation control units are controlled, and the dispatch instructions are updated every 60 minutes.

[0077] The present invention has the following advantages and effects compared with the prior art:

[0078] (1) The multi-scale, multi-group cognitive intelligent scheduling and control method for integrated energy systems proposed in this invention is divided into three different time scale control modes according to the control requirements of different generator sets. Furthermore, through the multi-time scale fusion framework, automatic generator sets and non-automatic generator sets are simultaneously scheduled and controlled to form an integrated integrated energy system scheduling and control method, which can avoid reverse scheduling.

[0079] (2) The multi-scale multi-group cognitive intelligent scheduling and control method for integrated energy systems proposed in this invention adopts an adaptive noise complete set empirical mode decomposition method in signal decomposition. Compared with other mode decomposition methods, this method still has good completeness with a smaller average number of iterations, and has a faster calculation speed and better mode decomposition results, which can make full preparations for the next step of signal prediction.

[0080] (3) The multi-scale, multi-group cognitive intelligent scheduling and control method for integrated energy systems proposed in this invention integrates human behavior cognitive knowledge matrix in long-time interval scheduling and control, considers user behavior and various climate factors, and can make scheduling instructions take into account all aspects, improve control accuracy and reduce frequency deviation. Attached Figure Description

[0081] Figure 1 This is a diagram illustrating the overall framework of power generation and dispatching in the method of this invention.

[0082] Figure 2 This is a flowchart of the bidirectional long short-term memory network prediction method of the present invention. Detailed Implementation

[0083] This invention proposes a multi-scale, multi-group cognitive intelligent scheduling and control method for integrated energy systems, which is described in detail below with reference to the accompanying drawings:

[0084] Figure 1 This is a diagram illustrating the overall framework of power generation and dispatching in this invention. First, it uses the frequency deviation sequence of the integrated energy system... As input, after adaptive noise complete set empirical mode decomposition, Decomposed into For each decomposed intrinsic mode vector and residual component, time series prediction is used for prediction. Based on the magnitude of the predicted signal fluctuations, it is categorized into large-fluctuation and small-fluctuation signals. For large-fluctuation and small-fluctuation signals, respectively, "state-action-reward-state-action" reinforcement learning and fractional-order "proportional-integral-derivative" methods are used for prediction and tracking. The two signals after prediction and tracking are superimposed as control commands to control the output of the automatic generator control unit every 4 seconds. Then, based on the region... Integrated energy system regional control error sequence As input, after adaptive noise complete set empirical mode decomposition and bidirectional long short-term memory network prediction, a consistency-based incremental rate method is used to output dispatch instructions with an interval of 5 minutes to the non-automatic generation control units. Finally, exist value of time User behavior information ,temperature ,humidity ,wind direction Wind speed ,precipitation and air pressure As matrix elements, they form an information matrix. The knowledge matrix of artificial behavior cognition that uses the alternating direction multiplier method for element updating Multiplying these together yields the non-automatic generator control unit scheduling instructions, which are updated every 60 minutes.

[0085] Figure 2 This is a flowchart of the bidirectional long short-term memory network prediction method of the present invention. First, in the forward channel long short-term memory neural network recurrent unit, the first decomposed signal... middle value of time and the external state at the previous moment As input, calculate the forget gate. Output at time Input gate Output at time Candidate status Output at time and output gate Output at time Then, in order to determine the internal state at the previous moment... The information deleted and added will change the internal state of the previous time step. With the Gate of Oblivion Output at time After multiplying corresponding elements, the result is compared with the candidate state. Output at time and input gate Output at time Add the results of multiplying corresponding elements together to get Internal state at time 1 Then, make Internal state at time 1 With output gate Output at time Multiplying corresponding elements yields external state at time , It also serves as a recurrent unit in a forward-channel long short-term memory neural network. The output at time step 1 also serves as part of the input for the next time step. Finally, in the reverse-channel long short-term memory neural network, in... The signal will be decomposed at any time. and the predicted signal from the previous moment As the input to the recurrent unit of the backward channel long short-term memory neural network, it undergoes the same operations as the recurrent unit of the forward channel long short-term memory neural network to obtain... external state at time , It also serves as a recurrent unit in a reverse-channel long short-term memory neural network. The output at any moment will and Adding them together yields a bidirectional long short-term memory neural network. Output at any given moment.

[0086] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1.A multi-scale multi-group cognitive intelligent scheduling and control method for integrated energy systems, characterized in that, The steps in use are: Step (1): In area A, obtain the frequency deviation sequence of the integrated energy system within a period of time and the integrated energy system area control error sequence between area and area , wherein ; obtain the behavior information of users at the moment , temperature , humidity , wind direction , wind speed , precipitation , and air pressure ; Step (2): Apply adaptive noise complete set empirical mode decomposition to the frequency deviation sequence. Process it, and set the number of iterations to 1. ,Will Positive and negative Gaussian white noise pairs are added to the original signal respectively. In total, for The first new signal, The new signal sequence after adding white noise. for: （1） wherein is the standard deviation of the noise; is the standard deviation of the noise; is the standard deviation of the noise; for each empirical mode decomposition, after the first empirical mode decomposition, the first mode component , the decomposed mode components are summed up and averaged, and the : （2） In the formula, is the 1st eigenmode component; Compute 1st residual component : （3） In addition Auxiliary noise signals after empirical mode decomposition of a pair of positive and negative Gaussian white noise: （4） In the formula, is the first new signal after adding the auxiliary noise for the first time; is a function for empirical mode decomposition. for each empirical mode decomposition, after the first empirical mode decomposition, the first mode component , the decomposed mode components are summed up and averaged, and the : （5） In the formula, is the 2nd eigenmode component; Computing the 2nd residual component : （6） Step (3): Repeat the calculation of step (2) until the residual component cannot be further decomposed, in which case the number of intrinsic mode components obtained after the next iteration is then the original signal is decomposed into: （7） wherein is the kth eigenmode component; is the kth residual component; Step (4): Put all and As input, a time series prediction method is used to process the input signal. The time series prediction neural network has two gating mechanisms in its recurrent unit: a reset gate and an update gate. In this prediction, middle value of time For input, reset the door. The output at time t is: （8） wherein, is the reset gate the output at time t; is the sigmoid function; is the state weight of the reset gate; is the input weight of the reset gate; is the bias of the reset gate; is the step size; is the hidden state of the cell at time t-1. Updating gate The output at the moment is: （9） wherein is an update gate the output at the time instant; is a state weight for the update gate and to update the input weights of the gate; to update the bias of the gate; Circulating body unit The activation state of the time implicit layer is: （10） wherein is a recurrent cell unit is the activation state of the hidden layer at time t; is an activation function; is a Hadamard product; is an input weight of the hidden layer; is a weight of the hidden layer; is a bias of the hidden layer; Computing a loop body element Output of the time instant: （11） wherein is a recurrent cell unit the output at time t, also serves as the hidden layer state at the next time t+1. Step (5): repeat the process of step (4) to get a total of predicted signals, denoted as the predicted signal , and classify them according to the fluctuation size, and in the same time , if the number of extreme points of the function image is greater than or equal to a set value , then it is classified as a large fluctuation signal; if the number of extreme points of the function image is less than a set value , then it is classified as a small fluctuation signal; for a large fluctuation signal, a "state-action-reward-state-action" learning method is used for reinforcement learning; for a small fluctuation signal, because the function changes slowly, a fractional order "proportional-integral-derivative" method is used for learning following; in the "state-action-reward-state-action" reinforcement learning method, the state , the action , and the reward expectation of the action （12） In the formula, For state Next action Expected rewards; The expected operation function; for Discount factor of time; for Momentary rewards; for The state at any given moment; for Actions at any given moment; To reinforce the learning state, i.e., the system frequency deviation ; This refers to the system's action, i.e., the unit's power output command; The processed signal of the large fluctuation signal is output after "state-action-reward-state-action" reinforcement learning ; In fractional order "proportional-integral-derivative" methods, the transfer function is: （13） wherein is a proportional regulation coefficient; is an integral regulation coefficient; is a derivative regulation coefficient; is an integral order; is a derivative order and ; The processing signal of the small fluctuation signal is: （14） wherein is the processed signal of the small fluctuation signal; is the inverse Laplace transform function; is the Laplace transform of the small fluctuation signal; Step (6): and The sum is used as the unit power output command, and the output of the automatic generation control unit is controlled in real time every 4 seconds. Step (7): Combine the results obtained in step (1) with the region Integrated energy system regional control error sequence As input, the decomposed signal is obtained through adaptive noise complete set empirical analysis. The decomposed signal is predicted using a bidirectional long short-term memory network; the recurrent unit of the long short-term memory neural network has three gated units: a forget gate, an input gate, and an output gate, using the first decomposed signal as an example. middle value of time For input, forget gate The output at time t is: （15） wherein is the forget gate is the output at time t; is the input weight of the forget gate is the state weight of the forget gate is the previous external state and output at time t-1 is the bias of the forget gate Input gate The output at time t is: （16） wherein is the input gate the output at time t; is the input weight of the input gate is the state weight of the input gate is the bias of the input gate candidate state The output at time t is: （17） wherein is the candidate state the output at time t; is the input weight for the candidate state is the state weight for the candidate state is the bias for the candidate state Output gate The output at the moment is: （18） wherein is the output gate the output at time t; is the input weight for the output gate is the state weight for the output gate is the bias for the output gate The internal state at time t is: （19） wherein is the internal state at the moment; is the internal state at the previous moment; Finally, The external states and outputs at time instant are: （20） wherein is the external state and output at the moment, while also serving as the input for the next moment; Step (8): repeat the process of step (7) to obtain the prediction signal in the form of The consistency increment rate criterion is applied to the prediction signal to obtain the scheduling instruction, and the scheduling instruction is updated every 5 minutes for the non-automatic generating unit. The output distribution formula of the consistency equal micro-increase rate criterion is: （21） wherein is the moment of the unit the load assigned; is the rate of increase of the consumption of the system at the moment; and is the consumption characteristic curve coefficient of the unit . Step (9): exist value of time User behavior information ,temperature ,humidity ,wind direction Wind speed ,precipitation and air pressure As matrix elements, they form an information matrix. ; （22） Constructing a matrix of artificial behavioral cognitive knowledge wherein is the number of non-automatic generation control units, and using an alternating direction multiplier method to update each parameter in the matrix of artificial behavioral cognitive knowledge . The information matrix is multiplied with the matrix of cognitive knowledge of human behavior to obtain the output matrix : （23） Output matrix Each element in the matrix The scheduling instruction is updated every 60 minutes as a dispatch instruction to control the corresponding non-automatic generating control unit.

Images