An adaptive economic dispatch system and method based on deep learning

Abstract

The present invention discloses an adaptive economic dispatch system based on deep learning, which is characterized by comprising a load matrix construction module, an adaptive learning model and training sample set construction module, a learning model training target construction module, a differentiated training sample set construction module, a long short-term memory network construction module and a real-time economic dispatch adaptive learning model construction module; the present invention generates a learning model training target based on the perfect dispatch concept, and trains the LSTM-IIU network pair using the massive historical data stored in the system. Once the learning model is constructed, the predicted data of the target dispatch moment is input to obtain the corresponding real-time economic dispatch plan.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and specifically to an adaptive economic dispatching system and method based on deep learning. Background Art

[0002] Economic dispatch is the core of economic operation of power system. In actual system operation, economic dispatch is regarded as a nonlinear optimization problem with uncertainty and is modeled and solved. The key is to accurately model the uncertainty factors such as system load forecast data and wind power output forecast data. Therefore, the accuracy of the dispatch plan mainly depends on the accuracy of modeling the uncertainty factors. However, with the continuous upgrading of load structure and the large-scale grid connection of renewable energy power generation, more uncertainties have been brought to the power system. The dispatch plan based on forecast data may seriously deviate from the actual dispatch plan. Once there is a large deviation between power generation and load demand, it may cause frequency deviation, which in turn affects the safe and stable operation of the system.

[0003] In order to effectively solve the deviation caused by uncertainty to the dispatching scheme, the existing economic dispatch modeling methods with uncertainty are usually divided into two categories, namely, probabilistic modeling methods and robust modeling methods. Among them, determining the probability density function of random variables such as load forecast value and new energy output forecast value is the core link of the probabilistic modeling method. It is generally believed that the load forecast value obeys the normal distribution, while the output forecast value of renewable energy such as wind power obeys the Weibull distribution, Beta distribution, etc. However, due to the inherent volatility, intermittency and randomness of renewable energy, it is difficult to accurately describe it with a fixed probability density function. Therefore, relevant scholars introduced fuzzy set theory into economic dispatch optimization modeling, and used fuzzy variables instead of random variables to achieve an accurate description of the forecast deviation of renewable energy output. In addition, some scholars introduced the opportunity constraint programming in the stochastic programming theory into the economic dispatch optimization modeling, that is, converting the constraints in the model into probability form. The dispatching scheme obtained by this method should meet the system constraints at a certain probability level, which directly reflects the dispatcher's consideration of the balance between system operation risk and operation cost. However, whether it is the selection of the membership function that characterizes the fuzzy variables in the fuzzy modeling method or the determination of the probability level in the chance-constrained programming modeling method, it is determined subjectively by the dispatching operator based on personal experience. Different dispatching operators may have different perceptions and judgments on the same scenario. In addition, the results obtained by probabilistic modeling are usually multiple or a group of dispatching schemes, and the dispatching operator needs to further screen to determine the best scheme. However, in the actual system operation, the dispatching operator needs an exact, objective, reliable and unique dispatching scheme. The core idea of ​​the economic dispatch solution method based on the idea of ​​robust optimization is to find extreme scenarios within the uncertainty range, and require that the dispatching scheme obtained in this scenario can still meet the system's operating constraints. Since this dispatching scheme considers the safety of system operation, it cannot meet the economic principle of dispatching to a certain extent.

[0004] With the rapid development of machine learning technology, relevant scholars have continued to use the basic idea of ​​economic dispatch as an optimization problem and tried to use reinforcement learning algorithms to model and solve it. The literature constructs a dynamic model of economic dispatch based on deep reinforcement learning, and uses a deep deterministic policy gradient algorithm to model the dispatch uncertainty caused by new energy. The literature attempts to apply the deep deterministic policy gradient algorithm to the dynamic economic dispatch of integrated energy systems. Although this type of method can effectively solve the uncertainty in the system, it essentially still regards the economic dispatch problem as an optimization problem and uses artificial intelligence technology as a means to solve the optimization problem, that is, the basic modeling idea has not changed.

[0005] In addition, in order to effectively solve the uncertainty in economic dispatch problems, the PJM power market in the United States proposed the concept of perfect dispatch. The core idea of ​​this method is to recalculate the output of each unit after the real-time market ends, based on a full understanding of the actual operating conditions of the system, such as the actual value of the load and the actual value of the output of new energy, and use it as the perfect dispatch plan for the real-time market. Compare the perfect dispatch plan with the real-time dispatch plan to find out the reasons for the deviation between the two dispatch plans, and provide experience guidance for future dispatch operations. However, at present, perfect dispatch is performed manually offline and cannot be executed automatically. Summary of the invention

[0006] The purpose of the present invention is to provide an adaptive economic dispatch system and method based on deep learning. The present invention generates a learning model training target based on the perfect dispatch concept, and uses the massive historical data stored in the system to train the LSTM-IIU network. Once the learning model is constructed, the predicted data of the target dispatch time is input to obtain the corresponding real-time economic dispatch plan.

[0007] To achieve this purpose, the present invention designs an adaptive economic dispatch system based on deep learning, which is characterized by: it includes a load matrix construction module, an adaptive learning model and training sample set construction module, a learning model training target construction module, a differentiated training sample set construction module, a long short-term memory network construction module and a real-time economic dispatch adaptive learning model construction module;

[0008] The load matrix construction module is used to construct a load matrix for each dispatching moment by utilizing the temporal continuity of the output of each unit in the load and the temporal and spatial complexity of the load forecast data;

[0009] The adaptive learning model and training sample set construction module is used to select the key historical moments of each scheduling moment by using matrix correlation analysis for the load prediction matrix for each scheduling moment, and determine the adaptive learning model for each scheduling moment based on the key historical moments of each scheduling moment;

[0010] The training target building module of the learning model is used to generate the training target of the adaptive learning model at each scheduling moment by using the perfect scheduling theory for the adaptive learning model at each scheduling moment;

[0011] The differentiated training sample set construction module is used to calculate the similarity coefficient between the load matrix at the current scheduling moment and the load matrix at the historical scheduling moment, and take the load matrix at the historical scheduling moment that is greater than the set similarity coefficient as the key feature and the feature that should be remembered of the load matrix at the current scheduling moment to form a differentiated sample set;

[0012] The long short-term memory network construction module is used to use the load forecast error of the adaptive learning model at each scheduling moment as the internal input of the long short-term memory network to obtain a long short-term memory network updated based on the load measurement error;

[0013] The real-time economic dispatch adaptive learning model construction module is used to construct a real-time economic dispatch adaptive learning model for each dispatching time using a long short-term memory network based on load measurement error update.

[0014] Beneficial effects of the present invention:

[0015] This paper proposes a new method for modeling and solving real-time economic dispatch problems based on deep learning. This method is based on the concept of perfect dispatch and uses the massive historical dispatch data stored in the system to train the LSTM-IIU (Long Short-term Memory based internal input update) network, and constructs an adaptive learning model for each dispatching moment to obtain the corresponding real-time economic dispatch plan.

[0016] The innovation of this method is mainly reflected in three aspects:

[0017] Firstly, this paper proposes a new approach to solve the real-time economic dispatch problem, which is regarded as a learning problem rather than an optimization problem, and uses the LSTM-IIU network to model and solve it; secondly, the spatiotemporal characteristics of the load are fully considered, and the matrix correlation analysis method is used to screen the key historical moments of each dispatching moment, and based on this, the adaptive structure of the learning model is determined and the corresponding differentiated training sample set is constructed; then, this paper takes the prediction error as the internal input of the LSTM network, proposes an LSTM-IIU network that is more suitable for the real-time economic dispatch problem, and uses this network to construct an adaptive learning model based on the LSTM-IIU network for different dispatching moments.

[0018] The adaptive learning model based on LSTM-IIU network for real-time economic dispatch proposed in this paper is a method to quickly extract knowledge from offline historical dispatch data. Once the model is built, the dispatcher can extract the corresponding model from the model library according to the target dispatch time, and input the prediction data at the corresponding time into the model to obtain the corresponding dispatch plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a structural schematic diagram of the present invention;

[0020] Figure 2 This is the correlation coefficient calculation process in the present invention;

[0021] Figure 3The LSTM network and its memory module structure in the present invention;

[0022] Figure 4 It is the LSTM-IIU network memory module structure in the present invention;

[0023] Among them, 1 is the load matrix construction module, 2 is the adaptive learning model and training sample set construction module, 3 is the training target construction module of the learning model, 4 is the differentiated training sample set construction module, 5 is the long short-term memory network construction module, 6 is the real-time economic dispatch adaptive learning model construction module, and 7 is the real-time economic dispatch adaptive learning model training module. DETAILED DESCRIPTION

[0024] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0025] like Figure 1 The deep learning-based adaptive economic dispatch system is characterized in that: it includes a load matrix construction module 1, an adaptive learning model and training sample set construction module 2, a learning model training target construction module 3, a differentiated training sample set construction module 4, a long short-term memory network construction module 5, a real-time economic dispatch adaptive learning model construction module 6 and a real-time economic dispatch adaptive learning model training module 7;

[0026] The load matrix construction module 1 is used to construct a load matrix for each scheduling moment by utilizing the temporal continuity of the output of each unit in the load and the temporal and spatial complexity of the load forecast data;

[0027] The adaptive learning model and training sample set construction module 2 is used to select the key historical moments of each scheduling moment by using matrix correlation analysis for the load prediction matrix for each scheduling moment, and determine the adaptive learning model for each scheduling moment based on the key historical moments of each scheduling moment;

[0028] The training target construction module 3 of the learning model is used to generate the training target of the adaptive learning model at each scheduling moment by using the perfect scheduling theory for the adaptive learning model at each scheduling moment;

[0029] The differentiated training sample set construction module 4 is used to calculate the similarity coefficient between the load matrix of the current scheduling moment and the load matrix of the historical scheduling moment, and use the load matrix of the historical scheduling moment that is greater than the set similarity coefficient as the key feature of the load matrix of the current scheduling moment to form a differentiated sample set. The key feature is used to improve the correlation between the current scheduling moment and the historical scheduling moment, thereby improving the learning and historical information memory capabilities of the model, so that the learning model has stronger adaptive capabilities;

[0030] The long short-term memory network construction module 5 is used to use the load forecast error of the adaptive learning model at each scheduling moment as the internal input of the long short-term memory network to obtain a long short-term memory network updated based on the load measurement error;

[0031] The real-time economic dispatch adaptive learning model construction module 6 is used to construct a real-time economic dispatch adaptive learning model for each dispatching moment using a long short-term memory network updated based on load measurement errors.

[0032] The real-time economic dispatch adaptive learning model training module 7 is used to input the differentiated training sample set and the training objectives of the adaptive learning model at each dispatching moment into the real-time economic dispatch adaptive learning model for each dispatching moment for training, and transmit the power grid section load to the real-time economic dispatch adaptive learning model at each dispatching moment after the training is completed to obtain the regulated output of each unit.

[0033] In the above technical scheme, the scheduling scheme has significant short-term continuity in time, that is, the output of each unit at the current moment is adjusted on the basis of the output of each unit at the previous moment. At the same time, due to the obvious differences in electricity demand at different scheduling moments, it is necessary to analyze the correlation between each scheduling moment and the historical moment to determine their respective key historical moments. Considering that the unit output is mainly affected by the system load, the correlation of the scheduling scheme in the time dimension can be characterized by the correlation of the system load. In addition, considering that there are differences in the load prediction errors of each load node, the load prediction value of each node is used instead of the total system load prediction value for correlation analysis. In order to fully reflect the spatiotemporal complexity of the load prediction data, the load prediction matrix for each scheduling moment is:

[0034]

[0035] In the formula, is the forecast load matrix corresponding to the dispatching time t, m is the time span of the load data, for example, when the sampling time span of the load data is 1 year, then m = 365, n is the number of load nodes in the system, represents the load forecast value at node n at the scheduling time t on the mth day;

[0036] The historical moment i of the dispatching time t ranges from 1 to k (the number of historical moments can be selected according to actual needs), and each dispatching moment has k forecast load matrices about the historical moments Its structure is as follows:

[0037]

[0038] In the formula, is the forecast load matrix at historical time ti, It represents the load forecast value at node n at historical time ti on the mth day;

[0039] To quantize the matrix and matrix The correlation between them is used to determine the key historical moments of each scheduling moment. The Pearson correlation coefficient method is extended from the vector level to the matrix level to calculate the correlation coefficient. The calculation formula is as follows:

[0040]

[0041] In the formula, r represents the correlation coefficient between the two matrices. and The matrices are and The average value of the elements in , ⊙ is matrix multiplication, ∑ means summing the elements in the matrix by row or column;

[0042] There are k historical moments in the dispatch time t, k is 96, and a load forecast matrix is ​​constructed for each historical moment Therefore, there are k correlation coefficients r about the scheduling time t. If the correlation coefficient r is greater than 0, it means and There is a positive correlation, and r The larger the value, the stronger the correlation between the two. The calculation process of the correlation coefficient r is as follows: Figure 1 As shown;

[0043] First, take any dispatch time t, then use formula 1 to construct the forecast load matrix for dispatch time t, select historical time i from 1 to 96, then use formula 2 to construct the forecast load matrix for historical time i, and then use formula 3 to calculate the correlation coefficient r;

[0044] After obtaining the correlation coefficient between the scheduling moment t and k historical moments, in order to determine its key historical moment, the correlation coefficients are arranged in descending order, and the order of magnitude of the difference between two adjacent correlation coefficients is calculated. The order of magnitude value and the learning effect of the network model are comprehensively considered from the two levels of physical meaning and data scale, and a suitable correlation coefficient is selected as the threshold. The historical moment corresponding to the correlation coefficient greater than the preset correlation coefficient threshold is the key historical moment of the scheduling moment. The key historical moment of each scheduling moment is used to determine the adaptive learning model for each scheduling moment.

[0045] The massive scheduling data stored in the system belongs to standard time series data, that is, the data not only has hidden time tags, but also is arranged in chronological order. Traditional neural networks can only analyze the relationship between the input and output of the same section, and cannot effectively process time series data. Therefore, relevant scholars have proposed a recurrent neural network (RNN) suitable for time series data analysis. RNN is a typical deep neural network, and its depth is reflected in the transmission and memory of data in the time dimension. Although RNN can consider the information transmission in time series data to a certain extent, it has the problem of gradient disappearance, that is, it cannot store and utilize long-term historical information. To solve this problem, relevant scholars have proposed an artificial neural network that uses memory function to enhance the influence of historical information on current output, namely LSTM network. The biggest difference between LSTM network and RNN in structure is that the internal composition of its hidden layer has been replaced by a memory module composed of gate structure instead of simple neurons. Figure 3 The structure of the LSTM network and its memory module is shown, where x t is the external input of the memory module at time t, i.e., the predicted value of each load node at time t; c t is the long-term memory corresponding to time t; h t-1 is the output of the LSTM network at time t-1, i.e., the predicted value of the output of each unit at time t-1, and is also input from the network to the memory module at time t as short-term memory; S(*) and T(*) represent the sigmoid activation function and the tanh activation function respectively;

[0046] Depend on Figure 3 It can be seen that in the LSTM network, the unit output h at the scheduling time t t Not only affected by the external input x at the same time t and is affected by the short-term memory h corresponding to the historical moment t-1 t-1 However, due to the short-term memory h t-1 The physical meaning is the predicted value of each unit output. The forget gate, input gate, and output gate need to control the forgetting of historical memory, the input of new information, and the output of the unit output at time t according to the predicted data. This update method cannot take into account the adjustment effect of the prediction error at time t-1 on the gate structure (that is, when the model is updated, the impact of the prediction result at the previous moment does not need to be considered). Once the short-term memory h is passed step by step with the time series t-1 When there is a large error between the actual value of the unit output and the actual value of the unit output, it is very easy to affect the final output of the model. At the same time, when the time series reaches the scheduling time t, the actual value of the unit output corresponding to the historical time t-1 is available, which lays a data foundation for the update of the gate structure using the prediction error control.

[0047] Therefore, based on the traditional LSTM network, the present invention proposes to introduce the actual value of the unit output corresponding to the historical moment into the internal input of the LSTM network, and use the prediction error between the unit output prediction value and the actual value obtained by the LSTM network itself as short-term memory to control the action of each gate structure, that is, the LSTM network is updated according to the internal input of the memory module. The internal structure of the LSTM-IIU network memory module proposed by the present invention is as follows: Figure 4 As shown. Among them, h′ t-1 h is short-term memory t-1 The corresponding actual value, e t-1 is the prediction error.

[0048] Depend on Figure 3 It can be seen that the gate structure of the LSTM-IIU network proposed in this paper has not changed compared with the traditional LSTM network. The internal input is obtained by introducing the actual value h′ that can be collected at the current moment. t-1 , and use the original predicted value h t-1 The prediction error obtained through a series of matrix calculations is input into the gate structure corresponding to time t, that is, the internal input of the long short-term memory network updated based on load measurement error at time t is:

[0049]

[0050] In the formula, e t-1 is the load forecast error, h′ t-1 h is short-term memory t-1 The actual value corresponding to represents the matrix addition operator;

[0051] In the long short-term memory network based on load measurement error update, the forget gate uses the unit output prediction error e at the previous time t-1 t-1 and the load forecast value x at the current time t t Adjust the forget control signal. If the prediction error is greater than the preset threshold, it means that e t-1 The prediction of the current time t is no longer effective, so this part of memory needs to be forgotten;

[0052] f t =sigmoid(W f ·[e t-1 ,x t ]+b f ) (5)

[0053] Where: W f and b f are the weight matrix and bias value corresponding to the forget gate, respectively, t Represents the gated output vector in the forget gate;

[0054] The input gate is divided into two parts. On the one hand, the prediction error e of the unit output at the previous time t-1 is used. t-1 and the load forecast value x at the current time t t , adjust the size of the input control signal, if the prediction error e t-1 If it is too large, you need to reduce the current memory information input to avoid affecting the prediction of future time series;

[0055] i t =sigmoid(W i ·[e t-1 ,x t ]+b i ) (6)

[0056] Where W i and b i are the weight matrix and bias value corresponding to the sigmoid activation function in the input gate, i t represents the gated output vector in the input gate;

[0057] On the other hand, using the unit output prediction error e at the previous time t-1 t-1 and the load forecast value x at the current time t t , generate the memory information corresponding to the current time t

[0058]

[0059] Where W c and b c are the weight matrix and bias value corresponding to the tanh activation function in the input gate, respectively. is the memory output vector in the input gate;

[0060] At this point, the long-term memory in the LSTM-IIU network has been composed of c t-1 Updated to c t ;

[0061]

[0062] In the formula, c t-1 Indicates the memory cell of the previous moment;

[0063] The output gate also uses e t-1 and x t The control adjusts the output control signal and determines the long-term memory c t How much is output?

[0064] o t =sigmoid(Wo ·[e t-1 ,x t ]+b o ) (9)

[0065] h t =o t ⊙tanh(c t ) (10)

[0066] Where: W o and b o are the weight matrix and bias value corresponding to the output gate, respectively. t represents the gated output vector in the output gate, h t Represents the hidden state of the neural network;

[0067] Formulas (4) to (10) constitute a long short-term memory network based on load measurement error update. It can be seen that the LSTM-IIU network proposed in the present invention introduces the actual value of the unit output, combines the predicted value output by the network, and uses the predicted error as the internal input of the memory module to adjust the control signal and memory update of each gate structure. Compared with the traditional LSMT network that uses the unit output predicted value generated by itself as the internal input, it is more in line with the actual business situation of economic dispatch.

[0068] In the above technical solution, the real-time economic dispatch adaptive learning model construction module 6 sets the number of memory modules in the long-short term memory network updated based on the load measurement error according to the number of key historical moments at different dispatching moments, and constructs a learning model with an adaptive number of memory modules for each dispatching moment.

[0069] As a deep learning model, the depth of the LSTM-IIU network is mainly reflected in the time level, that is, the depth of historical memory tracing. In the LSTM-IIU network, the number of memory modules directly reflects the model's ability to remember historical information. Therefore, it is necessary to set the number of memory modules in combination with the real-time economic dispatch business background.

[0070] Affected by load fluctuations, there are certain differences in the scheduling schemes corresponding to each scheduling moment. Therefore, the number of key historical moments obtained by matrix correlation analysis for each scheduling moment is different. If each scheduling moment shares a learning model, it is impossible to reflect its differences in historical information tracing. Therefore, the present invention uses the results of matrix correlation analysis to set the number of memory modules in each LSTM-IIU network according to the number of key historical moments at different scheduling moments, that is, to construct a learning model that is adaptive to the number of memory modules for different scheduling moments. For example, assuming that the number of key historical moments at the scheduling moment 00:00 is 10, the number of memory modules in the adaptive learning model corresponding to the scheduling moment is 11, of which the first 10 memory modules correspond to 10 historical key moments, which are used to transmit historical information, and the last memory module corresponds to the scheduling moment 00:00, which is used to output the model learning results. Since the number of key historical moments at different scheduling moments is different, the number of memory modules in the adaptive learning model for different scheduling moments is also different. The dispatch cycle of real-time economic dispatch is usually 15 minutes. There are 96 dispatch moments in a day, so a total of 96 adaptive learning models based on LSTM-IIU networks with different structures are obtained, that is, the structure of the learning model is adaptive.

[0071] In the adaptive learning model based on LSTM-IIU network, the input variables of the training sample set are composed of the load forecast value of the node and the output forecast value of each unit obtained based on the forecast value. Therefore, the dimension of the input variable is equal to the number of load nodes plus the number of generators. For example, if there are n lond load nodes and n gen generators, then the number of input independent variables is n x =n load +n gen . At the same time, since the dimensions of each input independent variable (load forecast value and unit output forecast value) are active power (MW) and the values ​​are of the same order of magnitude, the learning effect of the model will not be affected by the different dimensions, so there is no need to normalize the training data. The training target of the differentiated sample set is a perfect scheduling plan, that is, the output value of each unit calculated based on the actual load value collected using the Security Constrained Optimal Power Flow (SCOPF) model. The dimension of the training target is equal to the number of generators in the system minus 1, that is, n y =n gen -1;

[0072] The output of each unit in the training target of the differentiated sample set needs to meet the following conditions:

[0073] The optimization goal of the optimal power flow economic dispatch model is to minimize the power generation cost:

[0074]

[0075] Where: P g.i is the active power output of the i-th unit; a i , b i 、c i are the consumption coefficients of each item of the corresponding units respectively;

[0076] The equality constraint is the power balance equation:

[0077]

[0078] Where: P l is the total load of the adaptive economic dispatch system;

[0079] The inequality constraints include unit output constraints, unit ramp constraints, and line power constraints. The unit output constraints represent the active output range of the thermal power unit determined by the design:

[0080] P g.i.min ≤P g.i ≤P g.i.max (13)

[0081] Where: P g.i.max and P g.i.min They are respectively the upper and lower limits of active output of the corresponding units;

[0082] The unit ramp constraint represents the maximum active power that different thermal power units can adjust within the unit dispatching time:

[0083] r g.i.d T≤P g.i(t+1) -P g.i(t) ≤r g.i.u T (14)

[0084] Where: r g.i.d and r g.i.u They are the downward and upward ramp rates of the corresponding units per unit time, and T is the scheduling period;

[0085] Line power constraint, that is, it must not be greater than the maximum power that the line can pass;

[0086] S ij ≤S ij.max (15)

[0087] Where: S ij is the apparent power flowing between lines ij; S ij.max This is the maximum apparent power allowed for the line.

[0088] In the training process of the adaptive learning model based on the LSTM-IIU network, the training samples are input into the LSTM-IIU network in batches. The number of training samples in each batch is the same, and the product of the number of training samples and the number of batches is equal to the sum of the number of samples in the training set. The number of training samples in each batch is equal to the number of memory modules in the LSTM-IIU network, so as to ensure that each time a sample is input, there is only one training sample input into the memory module. The number of memory modules of the adaptive learning models for different scheduling moments is different, so the number of training samples corresponding to each learning model is also different, that is, the training sample set has differentiated characteristics in terms of both composition and scale. For example, taking the training samples with a time span of 1 year as an example, assuming that the number of key historical moments at the target scheduling moments 00:00 and 12:00 are 5 and 10 respectively, then in the adaptive learning models corresponding to the two scheduling moments, the number of memory modules is 6 and 11 respectively, and the number of training samples in the corresponding training sample set is 2190 (365*(5+1)) and 4015 (365*(10+1)) respectively.

[0089] In the above technical solution, the differentiated training sample set and the training target of the adaptive learning model at each dispatching moment are input into the real-time economic dispatch adaptive learning model for each dispatching moment for training, which includes two links: offline training and online application;

[0090] The offline training phase is to build an adaptive learning model for different target scheduling moments based on massive historical data. The system stores a large amount of historical operation data, such as the load forecast data stored in the prediction system and the historical scheduling plans stored in the Energy Management System (EMS). Based on the historical load forecast data, the key historical moments of each scheduling moment are determined by matrix correlation analysis, and then the internal structure of the long short-term memory network updated based on the load measurement error and the corresponding differentiated training sample set are determined;

[0091] At the same time, the perfect scheduling plan at the corresponding moment is used as the training target of the long short-term memory network based on the load measurement error update, and an adaptive learning model of the long short-term memory network based on the load measurement error update can be constructed. Once the model is constructed, it is stored in the offline model library, and the dispatcher can call it according to actual needs. Since the entire link only uses historical data, and the historical data is huge in scale and time-consuming to calculate, it only needs to be executed offline;

[0092] In the online application link, the dispatcher needs to extract the corresponding adaptive learning model according to the target dispatch time, and input the load forecast data obtained through the prediction system and the corresponding output of each unit into the real-time economic dispatch adaptive learning model for each dispatch time according to the training sample set structure, so as to obtain the output result of the real-time economic dispatch adaptive learning model for each dispatch time, which is the perfect dispatch plan corresponding to the target dispatch time. Since the current target dispatch time is the historical dispatch time of the future target dispatch time, the actual load data can be obtained at the end of the target dispatch time, and the perfect dispatch plan can be calculated using this data. In order to ensure the real-time update of the model, the perfect dispatch plan is passed back to the offline training link so that the offline model library can be updated in real time.

[0093] An adaptive economic dispatch method based on deep learning includes the following steps:

[0094] Step 1: Using the temporal continuity of the output of each unit in the load and the temporal and spatial complexity of the load forecast data, a load forecast matrix for each dispatching time is constructed;

[0095] Step 2: Matrix correlation analysis is used to select the key historical moments of each dispatching moment for the load forecast matrix at each dispatching moment, and the adaptive learning model for each dispatching moment is determined based on the key historical moments of each dispatching moment;

[0096] Step 3: Generate the training target of the adaptive learning model at each scheduling moment using the perfect scheduling theory for the adaptive learning model at each scheduling moment;

[0097] Step 4: Calculate the similarity coefficient between the load matrix at the current scheduling moment and the load matrix at the historical scheduling moment, and use the load matrix at the historical scheduling moment that is greater than the set similarity coefficient as the key feature and the feature that should be remembered of the load matrix at the current scheduling moment to form a differentiated sample set;

[0098] Step 5: The load forecast error of the adaptive learning model at each dispatching time (the difference between the real data and the data predicted by the model during the training process) is used as the internal input of the long short-term memory network to obtain a long short-term memory network updated based on the load measurement error;

[0099] Step 6: Use the long short-term memory network based on load measurement error update to build a real-time economic dispatch adaptive learning model for each dispatch time;

[0100] Step 7: Input the differentiated training sample set and the training objectives of the adaptive learning model at each dispatching moment into the real-time economic dispatch adaptive learning model for each dispatching moment for training, and transmit the grid section load to the real-time economic dispatch adaptive learning model at each dispatching moment after training to obtain the regulated output of each unit.

[0101] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in this field.

Claims

1. An adaptive economic dispatch system based on deep learning, characterized by: It includes a load matrix building module (1), an adaptive learning model and training sample set building module (2), a learning model training target building module (3), a differentiated training sample set building module (4), a long short-term memory network building module (5) and a real-time economic dispatch adaptive learning model building module (6); The load matrix construction module (1) is used to construct a load matrix for each dispatching time by utilizing the temporal continuity of the output of each unit in the load and the temporal and spatial complexity of the load forecast data; The adaptive learning model and training sample set construction module (2) is used to select the key historical moments of each scheduling moment by using matrix correlation analysis for the load prediction matrix for each scheduling moment, and determine the adaptive learning model for each scheduling moment based on the key historical moments of each scheduling moment; The training target building module (3) of the learning model is used to generate the training target of the adaptive learning model at each scheduling moment by using the perfect scheduling theory for the adaptive learning model at each scheduling moment; The differentiated training sample set construction module (4) is used to calculate the similarity coefficient between the load matrix at the current scheduling moment and the load matrix at the historical scheduling moment, and use the load matrix at the historical scheduling moment that is greater than the set similarity coefficient as the key feature of the load matrix at the current scheduling moment to form a differentiated sample set; The long short-term memory network construction module (5) is used to use the load forecasting error of the adaptive learning model at each scheduling moment as the internal input of the long short-term memory network to obtain a long short-term memory network updated based on the load measurement error; The real-time economic dispatch adaptive learning model construction module (6) is used to construct a real-time economic dispatch adaptive learning model for each dispatching time by using a long short-term memory network updated based on load measurement errors; The load prediction matrix for each dispatching time is: In the formula, is the forecast load matrix corresponding to the dispatching time t, m is the time span of the load data, n is the number of load nodes in the system, represents the load forecast value at node n at the scheduling time t on the mth day; The historical time i of the dispatch time t ranges from 1 to k, and each dispatch time has k forecast load matrices about the historical time Its structure is as follows: In the formula, is the forecast load matrix at historical time ti, It represents the load forecast value at node n at historical time ti on the mth day; To quantize the matrix and matrix The correlation between them is used to determine the key historical moments of each scheduling moment. The Pearson correlation coefficient method is extended from the vector level to the matrix level to calculate the correlation coefficient. The calculation formula is as follows: In the formula, r represents the correlation coefficient between the two matrices. and The matrices are and The average value of the elements in , ☉ is matrix multiplication, ∑ means summing the elements in the matrix by row or column; There are k historical moments in the dispatch time t, and a load forecast matrix is ​​constructed for each historical moment Therefore, there are k correlation coefficients r about the scheduling time t. If the correlation coefficient r is greater than 0, it means and There is a positive correlation, and the larger the r, the stronger the correlation between the two; After obtaining the correlation coefficient between the scheduling moment t and k historical moments, in order to determine its key historical moment, the correlation coefficients are arranged in descending order, and the order of magnitude of the difference between two adjacent correlation coefficients is calculated. The historical moment corresponding to the correlation coefficient greater than the preset correlation coefficient threshold is the key historical moment of the scheduling moment. The key historical moment of each scheduling moment is used to determine the adaptive learning model for each scheduling moment.

2. The deep learning-based adaptive economic dispatch system according to claim 1, characterized in that: It also includes a real-time economic dispatch adaptive learning model training module (7), which is used to input the differentiated training sample set and the training objectives of the adaptive learning model at each dispatching moment into the real-time economic dispatch adaptive learning model for each dispatching moment for training, and transmit the power grid section load to the real-time economic dispatch adaptive learning model at each dispatching moment after the training is completed to obtain the regulated output of each unit.

3. The deep learning-based adaptive economic dispatch system according to claim 1, characterized in that: The internal input of the LSTM network based on load error update at time t is: In the formula, e t-1 is the load forecast error, h′ t-1 h is short-term memory t-1 The actual value corresponding to represents the matrix addition operator; In the long short-term memory network based on load measurement error update, the forget gate uses the unit output prediction error e at the previous time t-1 t-1 and the load forecast value x at the current time t t Adjust the forget control signal. If the prediction error is greater than the preset threshold, it means that e t-1 The prediction of the current time t is no longer effective, so this part of memory needs to be forgotten; f t =sigmoid(W f ·[e t-1 ,x t ]+b f ) (5) Where: W f and b f are the weight matrix and bias value corresponding to the forget gate, respectively, t Represents the gated output vector in the forget gate; The input gate is divided into two parts. On the one hand, the prediction error e of the unit output at the previous time t-1 is used. t-1 and the load forecast value x at the current time t t , adjust the size of the input control signal; i t =sigmoid(W i ·[e t-1 ,x t ]+b i ) (6) Where W i and b i are the weight matrix and bias value corresponding to the sigmoid activation function in the input gate, i t represents the gated output vector in the input gate; On the other hand, using the unit output prediction error e at the previous time t-1 t-1 and the load forecast value x at the current time t t , generate the memory information corresponding to the current time t Where W c and b c are the weight matrix and bias value corresponding to the tanh activation function in the input gate, respectively. is the memory output vector in the input gate; At this point, the long-term memory in the LSTM-IIU network has been composed of c t-1 Updated to c t ; In the formula, c t-1 Indicates the memory cell of the previous moment; The output gate also uses e t-1 and x t The control adjusts the output control signal and determines the long-term memory c t How much output? o t =sigmoid(W o ·[e t-1 ,x t ]+b o ) (9) h t =o t ⊙tanh(c t ) (10) Where: W o and b o are the weight matrix and bias value corresponding to the output gate, respectively. t represents the gated output vector in the output gate, h t Represents the hidden state of the neural network; Formulas (4) to (10) constitute a long short-term memory network based on load measurement error update.

4. The deep learning-based adaptive economic dispatch system according to claim 3, characterized in that: The real-time economic dispatch adaptive learning model construction module (6) sets the number of memory modules in the long short-term memory network updated based on the load measurement error according to the number of key historical moments at different dispatch moments, and constructs a learning model with an adaptive number of memory modules for each dispatch moment.

5. The deep learning-based adaptive economic dispatch system according to claim 3, characterized in that: The differentiated training sample set and the training target of the adaptive learning model at each dispatching moment are input into the real-time economic dispatch adaptive learning model for each dispatching moment for training, including two links: offline training and online application; The offline training phase is to build an adaptive learning model for different target dispatching moments based on massive historical data. Based on historical load forecast data, matrix correlation analysis is used to determine the key historical moments of each dispatching moment, and then the internal structure of the long short-term memory network updated based on load measurement errors and the corresponding differentiated training sample set are determined; At the same time, the perfect scheduling plan at the corresponding moment is used as the training target of the long short-term memory network based on load measurement error update, so as to build an adaptive learning model of the long short-term memory network based on load measurement error update. Once the model is built, it is stored in the offline model library, and the dispatcher can call it according to actual needs. In the online application link, the dispatcher needs to extract the corresponding adaptive learning model according to the target dispatching time, and input the load forecast data obtained through the prediction system and the corresponding output of each unit into the real-time economic dispatching adaptive learning model for each dispatching time according to the training sample set structure. The output result of the real-time economic dispatching adaptive learning model for each dispatching time can be obtained, which is the perfect dispatching plan corresponding to the target dispatching time.

6. The deep learning-based adaptive economic dispatch system according to claim 3, characterized in that: The training target of the differentiated sample set is a perfect dispatching scheme, that is, the output value of each unit calculated by the safety-constrained optimal power flow economic dispatching model based on the collected actual load value, and the dimension of the training target is equal to the number of generators in the system minus 1; The output of each unit in the training target of the differentiated sample set needs to meet the following conditions: The optimization goal of the optimal power flow economic dispatch model is to minimize the power generation cost: Where: P g.i is the active power output of the i-th unit; a i , b i 、c i are the consumption coefficients of each item of the corresponding units respectively; The equality constraint is the power balance equation: Where: P l is the total load of the adaptive economic dispatch system; The inequality constraints include unit output constraints, unit ramp constraints, and line power constraints. The unit output constraints represent the active output range of the thermal power unit determined by the design: P g.i.min ￡P g.i ￡P g.i.max (13) Where: P g.i.max and P g.i.min They are respectively the upper and lower limits of active output of the corresponding units; The unit ramp constraint represents the maximum active power that different thermal power units can adjust within the unit dispatching time: r g.i.d T￡P g.i(t+1) -P g.i(t) ￡r g.i.u T (14) Where: r g.i.d and r g.i.u They are the downward and upward ramp rates of the corresponding units per unit time, and T is the scheduling period; Line power constraint, that is, it must not be greater than the maximum power that the line can pass; S ij ￡S ij.max (15) Where: S ij is the apparent power flowing between lines ij; S ij.max This is the maximum apparent power allowed for the line.

7. An adaptive economic dispatch method based on deep learning according to the system of claim 1, characterized in that: It includes the following steps: Step 1: Using the temporal continuity of the output of each unit in the load and the temporal and spatial complexity of the load forecast data, a load forecast matrix for each dispatching time is constructed; Step 2: Matrix correlation analysis is used to select the key historical moments of each dispatching moment for the load forecast matrix at each dispatching moment, and the adaptive learning model for each dispatching moment is determined based on the key historical moments of each dispatching moment; Step 3: Generate the training target of the adaptive learning model at each scheduling moment using the perfect scheduling theory for the adaptive learning model at each scheduling moment; Step 4: Calculate the similarity coefficient between the load matrix at the current scheduling moment and the load matrix at the historical scheduling moment, and use the load matrix at the historical scheduling moment that is greater than the set similarity coefficient as the key feature and the feature that should be remembered of the load matrix at the current scheduling moment to form a differentiated sample set; Step 5: Taking the load forecast error of the adaptive learning model at each dispatching moment as the internal input of the long short-term memory network to obtain a long short-term memory network updated based on the load measurement error; Step 6: Use the long short-term memory network based on load measurement error update to build a real-time economic dispatch adaptive learning model for each dispatch time; Step 7: Input the differentiated training sample set and the training objectives of the adaptive learning model at each dispatching moment into the real-time economic dispatch adaptive learning model for each dispatching moment for training, and transmit the grid section load to the real-time economic dispatch adaptive learning model at each dispatching moment after training to obtain the regulated output of each unit.

Images