Unit commitment decision-making method based on high-dimensional sample XOR sparse coding

Abstract

A unit commitment decision-making method based on high-dimensional sample exclusive or sparse coding comprises the following steps: step 1, performing coding preprocessing on a high-dimensional unit start-stop state variable matrix to obtain a low-dimensional unit start-stop state coding matrix; 2, the load matrix and the high-dimensional unit start-stop state variable matrix are paired to form an inventive high-dimensional unit commitment decision sample set, the high-dimensional unit start-stop state variable matrix is subjected to XOR sparse coding dimensionality reduction, and the load matrix and the low-dimensional unit start-stop state coding matrix are paired to form an inventive low-dimensional unit commitment decision sample set; 3, constructing an encoder-decoder network model suitable for the unit commitment decision, and training and solving the encoder-decoder network model suitable for the unit commitment decision by using the low-dimensional unit commitment decision sample set obtained in the step 2; the unit commitment decision is realized through the steps.

Description

Technical Field

[0001] The present invention relates to the field of power system and automation technology, and in particular to a unit combination decision method based on high-dimensional sample XOR sparse coding technology. Background Art

[0002] The unit commitment decision method driven by physical models can establish and solve mathematical models with different focuses for various practical application scenarios, but there are problems such as complex modeling and solving, difficult modification, insufficient applicability, and long research cycle, which cannot meet the needs of the rapid development of the power system. In contrast, the massive power generation data accumulated on the power supply side provides data support for system operation, making a series of research progress in data-driven unit commitment decision methods based on artificial intelligence and big data. However, most of these methods are based on supervised learning, with huge sample sizes and long training times, and the sample structure usually includes a load matrix and a unit start-stop state variable matrix.

[0003] The number of generators N and the dispatch cycle T in the power system are also important factors affecting the efficiency of the unit combination solution. If the sequence of unit start and stop state variables composed of 0-1 integer variables in the entire dispatch cycle is regarded as a matrix, the matrix exhibits a high-dimensional nature. However, this high-dimensional characteristic and large sample size pose considerable challenges to the algorithm's time and memory consumption.

[0004] Therefore, the applicant proposes a unit combination decision method based on high-dimensional sample XOR sparse coding technology to improve the efficiency of unit combination decision-making. Summary of the invention

[0005] The purpose of the present invention is to solve the technical problem of low unit operation efficiency due to high-dimensional characteristics and large sample size in the actual operation stage of the power grid and in the unit scheduling cycle, and to propose a unit combination decision technology based on high-dimensional sample XOR sparse coding technology.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A unit commitment decision method based on high-dimensional sample XOR sparse coding includes the following steps:

[0008] Step 1: Perform coding preprocessing on the high-dimensional unit start-stop state variable matrix to obtain a low-dimensional unit start-stop state coding matrix;

[0009] Step 2: Pair the load matrix with the high-dimensional unit start-stop state variable matrix to form a high-dimensional unit combination decision sample set of the invention, perform XOR sparse coding on the high-dimensional unit start-stop state variable matrix to reduce the dimension, and pair the load matrix with the low-dimensional unit start-stop state coding matrix to form a low-dimensional unit combination decision sample set of the invention;

[0010] Step 3: Construct an encoder-decoder network model suitable for unit commitment decision-making, and use the low-dimensional unit commitment decision sample set obtained in step 2 to train and solve the encoder-decoder network model of unit commitment decision-making;

[0011] The above steps are used to realize unit combination decision.

[0012] In step 1, the coding preprocessing of the high-dimensional unit start-stop state variable matrix includes the following sub-steps:

[0013] Step 1-1): Identify the k-th high-dimensional unit start-stop state variable matrix The dimensions, number of rows T and number of columns N;

[0014]

[0015] Where: is the k-th high-dimensional unit start-stop state variable matrix, where the subscript 0 represents the state that is not XOR-sparse and not encoded; is the start / stop state 0-1 variable of unit n at time t in the matrix. The value range of time t is t∈[1,T], and the value range of unit n is n∈[1,N]; the number of rows T is the time of unit combination decision; the number of columns N is the number of generating units;

[0016] Step 1-2): Take the i-th row in the k-th high-dimensional unit start-stop state variable matrix, i∈[1,T], vector As the reference row vector, perform XOR operation on each element of the remaining row vector bit by bit with the corresponding element of the i-th row;

[0017]

[0018]

[0019] Where: is the reference row vector composed of the start-stop state variables of N units in the k-th high-dimensional unit start-stop state variable matrix; To form the base row vector The elements represent the start / stop status 0-1 variables of unit n at time i; taking the tth row of the remaining rows as an example, then is the t-th row vector With the reference row vector The row vector after bit-by-bit XOR operation; is the exclusive OR operator;

[0020] Step 1-3): Compare all remaining row vectors with the reference row vector After performing XOR operation bit by bit, the elements contained in all row vectors can be combined into a new high-dimensional unit start-stop state variable matrix The subscript i indicates the state where the i-th row vector is XOR-sparse;

[0021]

[0022] Where: is a new high-dimensional unit start-stop state variable matrix obtained by performing XOR sparse operation bit by bit with the i-th row vector as the reference row;

[0023] The row vector after performing bit-by-bit XOR sparse operations Assign a decimal value to encode, assuming for but:

[0024]

[0025] Step 1-4): When the k-th high-dimensional unit start-stop state variable matrix The base row vector The row number i of the sample is completed once from 1 to T, and the XOR sparse coding cycle iteration is regarded as the high-dimensional unit start-stop state variable matrix completing a complete XOR sparse coding process, and the XOR sparse coding result Z of the sample is obtained. k ;

[0026]

[0027] Where: is the kth high-dimensional unit start-stop state variable matrix After a complete XOR sparse coding cycle, the k-th low-dimensional unit start-stop state coding matrix Z is obtained k ;

[0028] Step 1-5): If the XOR sparse coding results of two consecutive high-dimensional samples are the same, it is considered to have reached the end state and proceed to step 1-6); otherwise, go to step 1-2);

[0029] Step 1-6): Output the low-dimensional unit start-stop state encoding matrix Z k ;

[0030]

[0031] Where: z i The result is decimal encoding.

[0032] In step 2, a low-dimensional unit commitment decision sample set S is constructed low, the sample data includes: load matrix P and pre-processed low-dimensional unit start-stop state encoding matrix Z, including the following sub-steps:

[0033] Step 2-1): Based on the load data and unit dispatching plan collected by the monitoring equipment during the actual operation of the power grid, the load matrix P and the high-dimensional unit start-stop state variable matrix U are paired to form the high-dimensional unit combination decision sample set S of the invention. high , usually expressed as:

[0034]

[0035] Where: is the high-dimensional unit commitment decision sample set S high The kth load matrix of k is the high-dimensional unit commitment decision sample set S high The k-th high-dimensional unit start-stop state variable matrix, where U k The dimension is T×N; K is the total number of unit commitment decision samples in the sample set;

[0036] Step 2-2): Use the method in step 1 to reduce the dimension of the high-dimensional unit start-stop state variable matrix U by XOR sparse coding, and pair the load matrix P with the low-dimensional unit start-stop state coding matrix Z to form the invented low-dimensional unit combination decision sample set S low , usually expressed as:

[0037]

[0038] Where: Z k is the low-dimensional unit commitment decision sample set S high The k-th low-dimensional unit start-stop state encoding matrix, where Z k The dimension is T×1.

[0039] In step 3, an encoder-decoder network model suitable for unit commitment decision-making is constructed, and the low-dimensional unit commitment decision sample set S in step 2 is used low Training and solving the encoder-decoder network model for unit commitment decision-making includes the following sub-steps:

[0040] Step 3-1), construct an encoder-decoder network model suitable for unit combination decision-making, the input of the network model is the load matrix P of dimension T×1, and the output is the low-dimensional unit start-stop state encoding matrix Z of dimension T×1; wherein, the encoder network model of unit combination decision-making converts the load matrix P into a fixed-length vector expression C, and the decoder network model of unit combination decision-making converts the fixed-length vector expression C into a low-dimensional unit start-stop state encoding matrix Z, realizing the process from encoding to decoding of the unit combination sample matrix, and the main sub-steps are as follows:

[0041] Step 3-1-1): The encoder network model of unit combination decision reads the load matrix P according to the time step; at time t, the input features of the GRU neurons of the encoder network model include: the current input load P t and the hidden layer state h of the previous GRU neuron at time t-1 t-1 ; The hidden layer state h of the GRU neuron at time t can be obtained t , which is encoded into the intermediate state C by the encoder network model of the unit combination decision t Then input the next GRU neuron, specifically:

[0042] h t =f(h t-1 ,P t ) (10)

[0043] C t =h t (11)

[0044] Where: h t is the hidden layer state of the GRU neuron in the encoder network model for unit combination decision at time t; h t-1 is the hidden layer state of the previous GRU neuron at time t-1; P t is the load information at time t; f is the transformation function of the hidden layer of the recurrent neural network; C t is the intermediate state after being encoded by the encoder network model at time t;

[0045] Step 3-1-2): The intermediate state of the encoder network model of the unit commitment decision at time T is the concatenation of the hidden layer states of all time steps:

[0046] C T =q({h1,h2,…,h T}) (12)

[0047] Where: C T is the coded information output at time T; q is a custom nonlinear function;

[0048] Step 3-1-3): The decoder network model of unit commitment decision outputs the low-dimensional unit start-stop state encoding matrix Z according to the time step; the low-dimensional unit start-stop state encoding matrix Z at time t-1 is converted into t-1 and the hidden layer state s′ of the previous GRU neuron t-1 And the intermediate state C after being decoded by the decoder network model at time t t As the input of the GRU neuron of the decoder network model at time t, the hidden layer state s′ of the GRU neuron of the decoder network model at time t is output t-1 ;

[0049] s′ t =g(s′ t-1 ,Z t-1 ,C t ) (13)

[0050] C t =s′ t (14)

[0051] Where: s′ t The hidden layer state of the GRU neuron in the decoder network model for unit combination decision-making at time t; Z t-1 is the low-dimensional unit start-stop state encoding matrix output at time t-1; s′ t-1 is the hidden layer state of the previous GRU neuron at time t-1; C t is the intermediate state after being decoded by the decoder network model at time t; g is the transformation function of the hidden layer of the recurrent neural network;

[0052] Step 3-1-4): The probability information decoded and output at time T by the decoder network model of unit commitment decision is based on the set of low-dimensional unit start-stop state encoding matrices {Z1, Z2, …, Z T-1} and the intermediate state C after being decoded by the decoder network model at the previous time step T-1 Joint decision:

[0053]

[0054] Where: p(Z T ) is the probability that the decoder outputs the decoded information at the last time step T; {Z1, Z2, …, Z T-1 The set C is the set of low-dimensional unit start-stop state encoding matrices output at the previous T-1 time; T-1 is the intermediate state after being decoded by the decoder at the previous time step;

[0055] Step 3-2), using the low-dimensional unit commitment decision sample set S in step 2 low The encoder-decoder network model suitable for unit commitment decision established in training step 3-1) can obtain the general mapping relationship of unit commitment decision:

[0056] (Z) = p(P L ) (16) Where: P L is the load matrix; Z is the low-dimensional unit start-stop state encoding matrix; p(*) is the load matrix P L The probability mapping model with the low-dimensional unit start-stop state encoding matrix Z;

[0057] The trained encoder-decoder network model for unit commitment decision-making can learn the mapping between the daily load matrix and the unit start-stop state variable encoding matrix in an offline state;

[0058] Step 3-3), input the daily load matrix and the unit start-stop state variable matrix to the optimal power flow (OPF) model, and solve the output power of the generator based on the optimal power flow (OPF) model, including the following sub-steps:

[0059] Step 3-3-1): The objective function of the OPF model is to minimize the operating cost of the unit:

[0060]

[0061] Where: F(P G ) is the operating cost of a conventional generator set; a i ,b i ,c i are the coal consumption coefficients of conventional power generating units; P Git is the output power of conventional generator set i at time t, U Git is the start and stop status of conventional generator set i at time t;

[0062] Step 3-3-2): Decision variable P Git Power balance constraints, output upper and lower limit constraints, climbing / slip rate constraints, and network security constraints must also be met to ensure the safe operation of the system:

[0063]

[0064]

[0065]

[0066]

[0067] Where: P Lt is the daily load at time t; and are the lower and upper limits of active output of unit i respectively; and are the sliding rate and climbing rate of unit i respectively; P lmax is the maximum transmission capacity allowed by line l in the base state; G l-i is the generator output power transfer distribution factor of unit i to line l; B is the number of nodes in the power grid; D bt is the load of node b in the power grid during period t;

[0068] Step 3-3-3): Output the active power of the unit according to the OPF model.

[0069] An encoder-decoder network model for unit combination decision-making, comprising an encoder network model for unit combination decision-making and a decoder network model for unit combination decision-making; the input of the encoder-decoder network model for unit combination decision-making is a load matrix P with a dimension of T×1, and the output is a low-dimensional unit start-stop state encoding matrix Z with a dimension of T×1; wherein the encoder network model for unit combination decision-making converts the load matrix P into a fixed-length vector expression C, and the decoder network model for unit combination decision-making converts the fixed-length vector expression C into a low-dimensional unit start-stop state encoding matrix Z, so as to realize the process from encoding to decoding of the unit combination sample matrix.

[0070] The encoder-decoder network model for unit commitment decision-making, when working, includes the following steps:

[0071] Step S1): The encoder network model for unit commitment decision reads the load matrix P according to the time step;

[0072] Specifically, at time t, the input features of the GRU neurons in the encoder network model include: the current input load P t and the hidden layer state h of the previous GRU neuron at time t-1 t-1 ; The hidden layer state h of the GRU neuron at time t can be obtained t , which is encoded into the intermediate state C by the encoder network model of the unit combination decision t Then input the next GRU neuron, which is specifically expressed as:

[0073] h t =f(h t-1 ,P t ) (10)

[0074] C t =h t (11)

[0075] Where: h t is the hidden layer state of the GRU neuron in the encoder network model for unit combination decision at time t; h t-1 is the hidden layer state of the previous GRU neuron at time t-1; P t is the load information at time t; f is the transformation function of the hidden layer of the recurrent neural network; C t is the intermediate state after being encoded by the encoder network model at time t;

[0076] Step S2): The encoder network model of the unit commitment decision encodes and outputs the intermediate state at time T, and the intermediate state is the concatenation of the hidden layer states of all time steps:

[0077] C T =q({h1,h2,…,h T}) (12)

[0078] Where: C T is the coded information output at time T; q is a custom nonlinear function;

[0079] Step S3): the decoder network model of unit commitment decision outputs a low-dimensional unit start-stop state encoding matrix Z according to the time step;

[0080] Specifically: the low-dimensional unit start-stop state encoding matrix Z at time t-1 t-1 and the hidden layer state s′ of the previous GRU neuron t-1 And the intermediate state C after being decoded by the decoder network model at time t t As the input of the GRU neuron of the decoder network model at time t, the hidden layer state s′ of the GRU neuron of the decoder network model at time t is output t-1 ;

[0081] s′ t =g(s′ t-1 ,Z t-1 ,C t ) (13)

[0082] C t =s′ t (14)

[0083] Where: s′ t The hidden layer state of the GRU neuron in the decoder network model for unit combination decision at time t; Z t-1 is the low-dimensional unit start-stop state encoding matrix output at time t-1; s′ t-1 is the hidden layer state of the previous GRU neuron at time t-1; C t is the intermediate state after being decoded by the decoder network model at time t; g is the transformation function of the hidden layer of the recurrent neural network;

[0084] Step S4): The decoder network model of the unit commitment decision decodes and outputs probability information at time T, wherein the probability information is based on the low-dimensional unit start-stop state encoding matrix set {Z1, Z2, ..., Z T-1} and the intermediate state C after being decoded by the decoder network model at the previous time step T-1 Joint decision:

[0085]

[0086] Where: p(Z T) is the probability that the decoder outputs the decoded information at the last time step T; {Z1, Z2, …, Z T-1 The set C is the set of low-dimensional unit start-stop state encoding matrices output at the previous T-1 time; T-1 It is the intermediate state after being decoded by the decoder at the previous time step.

[0087] The trained encoder-decoder network model for unit combination decision-making can learn the mapping between the daily load matrix and the unit start-stop state variable encoding matrix in an offline state, so that other models (such as the OPF model when solving the unit's output power) can obtain the unit's output power by inputting the daily load matrix and the unit start-stop state variable matrix.

[0088] Compared with the prior art, the present invention has the following technical effects:

[0089] 1) The present invention utilizes a high-dimensional sample XOR sparse coding preprocessing method to compress a high-dimensional unit combination start-stop state variable matrix U with a dimension of T×N into a low-dimensional unit combination start-stop state coding matrix Z with a dimension of T×1. This technology can effectively improve the efficiency of unit combination decision-making;

[0090] 2) On the one hand, the present invention can use the XOR sparsely coded unit start and stop status samples to offline train the unit combination decision deep learning model, which can reduce the model training time from 830s to 479s, and improve the training efficiency by 42.29%. On the other hand, the XOR sparse coding time of the high-dimensional unit start and stop status samples is only 0.1266s;

[0091] 3) Compared with traditional deep learning models, the unit combination decision encoder-decoder deep learning model used for offline training and online optimization in the present invention can achieve optimal decision-making of the model without clustering preprocessing samples to train multiple class models. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0093] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0094] Figure 2 This is an operation mode diagram of high-dimensional sample XOR sparse encoding and decoding in the present invention;

[0095] Figure 3 The encoder-decoder network model diagram of the unit combination decision in the present invention;

[0096] Figure 4 This is a schematic diagram of the comparison of the day-ahead planned time of the unit of the No. 91 test sample in the embodiment;

[0097] Figure 5 This is a schematic diagram of the comparison of the day-ahead planned time of the unit of the No. 92 test sample in the embodiment;

[0098] Figure 6 This is a schematic diagram of the comparison of the day-ahead planned time of the unit of the No. 93 test sample in the embodiment. DETAILED DESCRIPTION

[0099] like Figure 1 As shown, a unit commitment decision method based on high-dimensional sample XOR sparse coding includes the following steps:

[0100] Step 1: Perform coding preprocessing on the high-dimensional unit start-stop state variable matrix to obtain a low-dimensional unit start-stop state coding matrix;

[0101] Step 2: Pair the load matrix with the high-dimensional unit start-stop state variable matrix to form a high-dimensional unit combination decision sample set of the invention, perform XOR sparse coding on the high-dimensional unit start-stop state variable matrix to reduce the dimension, and pair the load matrix with the low-dimensional unit start-stop state coding matrix to form a low-dimensional unit combination decision sample set of the invention;

[0102] Step 3: Construct an encoder-decoder network model suitable for unit commitment decision-making, and use the low-dimensional unit commitment decision sample set obtained in step 2 to train and solve the encoder-decoder network model of unit commitment decision-making;

[0103] The above steps are used to realize unit combination decision.

[0104] In step 1, the encoding preprocessing process of the high-dimensional unit start-stop state variable matrix is ​​as follows: Figure 2 As shown, it includes the following sub-steps:

[0105] Step 1-1): Identify the k-th high-dimensional unit start-stop state variable matrix The dimensions, number of rows T and number of columns N;

[0106]

[0107] Where: is the k-th high-dimensional unit start-stop state variable matrix, where the subscript 0 represents the state that is not XOR-sparse and not encoded; is the start / stop state 0-1 variable of unit n at time t in the matrix. The value range of time t is t∈[1,T], and the value range of unit n is n∈[1,N]. The number of rows T is the time of unit combination decision; the number of columns N is the number of generating units.

[0108] Step 1-2): Take the i-th row in the k-th high-dimensional unit start-stop state variable matrix, i∈[1,T], vector As the reference row vector, perform XOR operation on each element of the remaining row vector bit by bit with the corresponding element of the i-th row;

[0109]

[0110]

[0111] Where: is the reference row vector composed of the start-stop state variables of N units in the k-th high-dimensional unit start-stop state variable matrix; To form the base row vector The elements represent the start / stop status 0-1 variables of unit n at time i; taking the tth row of the remaining rows as an example, then is the t-th row vector With the reference row vector The row vector after bit-by-bit XOR operation; Is the exclusive OR operator.

[0112] Step 1-3): Compare all remaining row vectors with the reference row vector After performing XOR operation bit by bit, the elements contained in all row vectors can be combined into a new high-dimensional unit start-stop state variable matrix The subscript i indicates the state where the i-th row vector is XOR-sparse;

[0113]

[0114] Where: It is a new high-dimensional unit start-stop state variable matrix obtained by performing XOR sparse operation bit by bit with the i-th row vector as the reference row.

[0115] The row vector after performing bit-by-bit XOR sparse operations Assign a decimal value to encode, assuming for but:

[0116]

[0117] Step 1-4): When the k-th high-dimensional unit start-stop state variable matrix The base row vector The row number i of the sample is completed once from 1 to T, and the XOR sparse coding cycle iteration is regarded as the high-dimensional unit start-stop state variable matrix completing a complete XOR sparse coding process, and the XOR sparse coding result Z of the sample is obtained. k ;

[0118]

[0119] Where: is the kth high-dimensional unit start-stop state variable matrix After a complete XOR sparse coding cycle, the k-th low-dimensional unit start-stop state coding matrix Z is obtained k .

[0120] Step 1-5): If the XOR sparse coding results of two consecutive high-dimensional samples are the same, it is considered to have reached the end state and proceed to step 1-6); otherwise, go to step 1-2);

[0121] Step 1-6): Output the low-dimensional unit start-stop state encoding matrix Z k .

[0122]

[0123] Where: z i The result is decimal encoding.

[0124] In step 2, a low-dimensional unit commitment decision sample set S is constructed low , the sample data includes: load matrix P and pre-processed low-dimensional unit start-stop state encoding matrix Z, including the following sub-steps:

[0125] Step 2-1): Based on the load data and unit dispatching plan collected by the monitoring equipment during the actual operation of the power grid, the load matrix P and the high-dimensional unit start-stop state variable matrix U are paired to form the high-dimensional unit combination decision sample set S of the invention. high , usually expressed as:

[0126]

[0127] Where: is the high-dimensional unit commitment decision sample set S high The kth load matrix of k is the high-dimensional unit commitment decision sample set S high The k-th high-dimensional unit start-stop state variable matrix, where U k The dimension is T×N; K is the total number of unit commitment decision samples in the sample set.

[0128] Step 2-2): Use the method in step 1 to reduce the dimension of the high-dimensional unit start-stop state variable matrix U by XOR sparse coding, and pair the load matrix P with the low-dimensional unit start-stop state coding matrix Z to form the invented low-dimensional unit combination decision sample set S low , usually expressed as:

[0129]

[0130] Where: Z k is the low-dimensional unit commitment decision sample set S highThe k-th low-dimensional unit start-stop state encoding matrix, where Z k The dimension is T×1.

[0131] In step 3, an encoder-decoder network model suitable for unit commitment decision-making is constructed, and the low-dimensional unit commitment decision sample set S in step 2 is used low Training and solving the encoder-decoder network model for unit commitment decision-making includes the following sub-steps:

[0132] Step 3-1), using the gated recurrent unit (GRU) as the basic unit and combining the basic architecture of the encoder-decoder, a encoder-decoder network model suitable for unit combination decision-making is constructed. The encoder-decoder network model for unit combination decision-making is shown in Figure 3 As shown in Figure 1, the input of the network model is a load matrix P with a dimension of T×1, and the output is a low-dimensional unit start-stop state encoding matrix Z with a dimension of T×1. Among them, the encoder network model of the unit combination decision converts the load matrix P into a fixed-length vector expression C, and the decoder network model of the unit combination decision converts the fixed-length vector expression C into a low-dimensional unit start-stop state encoding matrix Z, realizing the process from encoding to decoding of the unit combination sample matrix. The main sub-steps are as follows:

[0133] Step 3-1-1): The encoder network model for unit commitment decision reads the load matrix based on the time step At this moment, the input features of the GRU neuron in the encoder network model include: the current input load P t and the hidden layer state h of the previous GRU neuron at time t-1 t-1 The hidden layer state h of the GRU neuron at time t can be calculated t , which is encoded into the intermediate state C by the encoder network model of the unit combination decision t Then input the next GRU neuron.

[0134] h t =f(h t-1 ,P t ) (10)

[0135] C t =h t (11)

[0136] Where: h t is the hidden layer state of the GRU neuron in the encoder network model for unit combination decision at time t; h t-1 is the hidden layer state of the previous GRU neuron at time t-1; P t is the load information at time t; f is the transformation function of the hidden layer of the recurrent neural network; Ct It is the intermediate state after being encoded by the encoder network model at time t.

[0137] Step 3-1-2): The intermediate state of the encoder network model of the unit commitment decision at time T is the concatenation of the hidden layer states of all time steps:

[0138] C T =q({h1,h2,…,h T}) (12)

[0139] Where: C T is the coded information output at time T; q is a custom nonlinear function.

[0140] Step 3-1-3): The decoder network model of unit commitment decision outputs the low-dimensional unit start-stop state encoding matrix Z according to the time step. The low-dimensional unit start-stop state encoding matrix Z at time t-1 is t-1 and the hidden layer state s of the previous GRU neuron t ' -1 And the intermediate state C after being decoded by the decoder network model at time t t As the input of the GRU neuron of the decoder network model at time t, the hidden layer state s′ of the GRU neuron of the decoder network model at time t is output t-1 ;

[0141] s′ t =g(s′ t-1 ,Z t-1 ,C t ) (13)

[0142] C t =s′ t (14)

[0143] Where: s′ t The hidden layer state of the GRU neuron in the decoder network model for unit combination decision-making at time t; Z t-1 is the low-dimensional unit start-stop state encoding matrix output at time t-1; s′ t-1 is the hidden layer state of the previous GRU neuron at time t-1; C t is the intermediate state after being decoded by the decoder network model at time t; g is the transformation function of the hidden layer of the recurrent neural network;

[0144] Step 3-1-4): The probability information decoded and output at time T by the decoder network model of unit commitment decision is based on the set of low-dimensional unit start-stop state encoding matrices {Z1, Z2, …, Z T-1} and the intermediate state C after being decoded by the decoder network model at the previous time step T-1Joint decision:

[0145]

[0146] Where: p(Z T ) is the probability that the decoder outputs the decoded information at the last time step T; {Z1, Z2, …, Z T-1 The set C is the set of low-dimensional unit start-stop state encoding matrices output at the previous T-1 time; T-1 It is the intermediate state after being decoded by the decoder at the previous time step.

[0147] Step 3-2), using the low-dimensional unit commitment decision sample set S in step 2 low The encoder-decoder network model suitable for unit commitment decision established in training step 3-1) can obtain the general mapping relationship of unit commitment decision:

[0148] (Z) = p(P L ) (16)

[0149] Where: P L is the load matrix; Z is the low-dimensional unit start-stop state encoding matrix; p(*) is the load matrix P L Probabilistic mapping model with the low-dimensional unit start-stop state encoding matrix Z.

[0150] Step 3-3) is to solve the output power of the generator based on the acquired unit start-stop state variable matrix based on the optimal power flow (OPF) model, including the following sub-steps:

[0151] Step 3-3-1): The objective function of the OPF model is to minimize the operating cost of the unit:

[0152]

[0153] Where: F(P G ) is the operating cost of a conventional generator set; a i ,b i ,c i are the coal consumption coefficients of conventional power generating units; P Git is the output power of conventional generator set i at time t, U Git is the start and stop status of conventional generator set i at time t.

[0154] Step 3-3-2): Decision variable P Git Power balance constraints, output upper and lower limit constraints, climbing / slip rate constraints, and network security constraints must also be met to ensure the safe operation of the system:

[0155]

[0156]

[0157]

[0158]

[0159] Where: P Lt is the daily load at time t; and are the lower and upper limits of active output of unit i respectively; and are the sliding rate and climbing rate of unit i respectively; P lmax is the maximum transmission capacity allowed by line l in the base state; G l-i is the generator output power transfer distribution factor of unit i to line l; B is the number of nodes in the power grid; D bt is the load of node b in the power grid during period t.

[0160] Step 3-3-3): Output the active power of the unit according to the OPF model.

[0161] The paper "Research on Intelligent Decision-making Method for Unit Combination Based on Data Driven Self-learning Ability" proposed a unit combination decision-making technology based on long short-term memory network, which established the mapping between system load and dispatch decision results through historical data training, and made unit combination decisions based on this. However, on the one hand, this technology requires clustering preprocessing of historical dispatch data, and establishing mapping decisions based on long short-term memory network for different load types; this leads to low efficiency of unit combination decision-making.

[0162] Therefore, in order to solve the above technical problems, the present invention also proposes an encoder-decoder network model for unit combination decision-making, which includes an encoder network model for unit combination decision-making and a decoder network model for unit combination decision-making; the input of the encoder-decoder network model for unit combination decision-making is a load matrix P with a dimension of T×1, and the output is a low-dimensional unit start-stop state encoding matrix Z with a dimension of T×1; wherein, the encoder network model for unit combination decision-making converts the load matrix P into a fixed-length vector expression C, and the decoder network model for unit combination decision-making converts the fixed-length vector expression C into a low-dimensional unit start-stop state encoding matrix Z, so as to realize the process from encoding to decoding of the unit combination sample matrix.

[0163] The encoder-decoder network model for unit commitment decision-making, when working, includes the following steps:

[0164] Step S1): The encoder network model for unit commitment decision reads the load matrix P according to the time step;

[0165] Specifically, at time t, the input features of the GRU neurons in the encoder network model include: the current input load P t and the hidden layer state h of the previous GRU neuron at time t-1 t-1 ; The hidden layer state h of the GRU neuron at time t can be obtained t , which is encoded into the intermediate state C by the encoder network model of the unit combination decision t Then input the next GRU neuron, which is specifically expressed as:

[0166] h t =f(h t-1 ,P t ) (10)

[0167] C t =h t (11)

[0168] Where: h t is the hidden layer state of the GRU neuron in the encoder network model for unit combination decision at time t; h t-1 is the hidden layer state of the previous GRU neuron at time t-1; P t is the load information at time t; f is the transformation function of the hidden layer of the recurrent neural network; C t is the intermediate state after being encoded by the encoder network model at time t;

[0169] Step S2): The encoder network model of the unit commitment decision encodes and outputs the intermediate state at time T, and the intermediate state is the concatenation of the hidden layer states of all time steps:

[0170] C T =q({h1,h2,…,h T}) (12)

[0171] Where: C T is the coded information output at time T; q is a custom nonlinear function;

[0172] Step S3): the decoder network model of unit commitment decision outputs a low-dimensional unit start-stop state encoding matrix Z according to the time step;

[0173] Specifically: the low-dimensional unit start-stop state encoding matrix Z at time t-1 t-1 and the hidden layer state s′ of the previous GRU neuron t-1 And the intermediate state C after being decoded by the decoder network model at time t t As the input of the GRU neuron of the decoder network model at time t, the hidden layer state s′ of the GRU neuron of the decoder network model at time t is output t-1 ;

[0174] s′ t =g(s′ t-1 ,Z t-1 ,C t ) (13)

[0175] C t =s′ t (14)

[0176] Where: s′ t The hidden layer state of the GRU neuron in the decoder network model for unit combination decision-making at time t; Z t-1 is the low-dimensional unit start-stop state encoding matrix output at time t-1; s′ t-1 is the hidden layer state of the previous GRU neuron at time t-1; C t is the intermediate state after being decoded by the decoder network model at time t; g is the transformation function of the hidden layer of the recurrent neural network;

[0177] Step S4): The decoder network model of the unit commitment decision decodes and outputs probability information at time T, wherein the probability information is based on the low-dimensional unit start-stop state encoding matrix set {Z1, Z2, ..., Z T-1} and the intermediate state C after being decoded by the decoder network model at the previous time step T-1 Joint decision:

[0178]

[0179] Where: p(Z T ) is the probability that the decoder outputs the decoded information at the last time step T; {Z1, Z2, …, Z T-1 The set C is the set of low-dimensional unit start-stop state encoding matrices output at the previous T-1 time; T-1 It is the intermediate state after being decoded by the decoder at the previous time step.

[0180] The trained encoder-decoder network model for unit combination decision-making can learn the mapping between the daily load matrix and the unit start-stop state variable encoding matrix in an offline state, so that other models (such as the OPF model when solving the unit's output power) can obtain the unit's output power by inputting the daily load matrix and the unit start-stop state variable matrix.

[0181] The encoder-decoder network model for unit commitment decision-making proposed in the present invention utilizes the encoder-decoder network architecture to avoid clustering of historical load matrices, thereby simplifying the implementation process of the technology.

[0182] Example:

[0183] The present invention uses the IEEE 118-node system as an example for simulation verification. The load data is derived from the actual operation data of the Hunan power grid. After cleaning and screening, 93 groups of load matrices of typical days are obtained. In order to facilitate subsequent calculations and analysis, the k-means clustering algorithm is used to divide the above load matrix into three load scenarios A, B, and C. The training and test division of the sample set is shown in Table 1.

[0184] Table 1 Division of sample set

[0185]

[0186] Specifically, in order to comprehensively examine the advantages of the method of the present invention, the following comparative simulations will be performed:

[0187] Method 1: Data-driven unit commitment decision method without considering XOR sparse coding.

[0188] Method 2: Data-driven unit commitment decision method based on XOR sparse coding strategy, i.e., the method of the present invention

[0189] The deep learning models used in the above methods are all built based on the TensorFlow framework, in which the recurrent layer uses GRU neurons. The proposed model is executed on a workstation equipped with an eight-thread 3.0GHz CPU and 16.0GB RAM memory using Python 3.8.16, and the program compiler is PyCharm 2020.

[0190] (1) Analysis of the impact of XOR sparse coding method on decision results

[0191] In order to verify the correctness of the method of the present invention, the unit combination decision method based on Benders decomposition is used to solve the high-dimensional unit start-stop state variable matrix of the No. 1-No. 93 load matrix, and the XOR sparse coding method for high-dimensional samples proposed in the present invention is used to encode the No. 1-No. 93 high-dimensional unit start-stop state variable matrix into a decimal low-dimensional unit start-stop state coding matrix. The coding results are shown in Table 2.

[0192] Table 2 Coding results of the start-stop schemes for sample units No. 1-No. 93

[0193]

[0194] As can be seen from Table 2, for those samples with similar daily load matrices, the corresponding low-dimensional unit start-stop state encoding matrices are the same. The reason for this phenomenon is that the high-dimensional unit start-stop state variable matrix obtained by solving the load matrix with similar daily load using the unit combination decision method based on Benders decomposition is also the same. Therefore, a low-dimensional unit start-stop state encoding matrix may have a mapping relationship with multiple similar load matrices. The core work of the present invention is to hope that the machine can quickly identify the low-dimensional unit start-stop state encoding matrix mapped by the load matrix, and the reason for the different matrix encoding results is caused by the difference of the load matrix, but this does not affect the encoder-decoder network model applicable to unit combination decision-making to learn the mapping relationship between the load matrix and the low-dimensional unit start-stop state encoding matrix in an offline state. In summary, when executing the XOR sparse encoding step of the high-dimensional unit start-stop state variable matrix, it is not necessary to clearly distinguish the load matrix corresponding to the low-dimensional unit start-stop state encoding matrix, and similar load matrix samples are exactly the sample sets required by the encoder-decoder network model of the unit combination decision.

[0195] The No.1-No.90 load matrices are paired with the corresponding high-dimensional unit start-stop state variable matrices to form a low-dimensional unit commitment decision sample set for this example. After the encoder-decoder network model of the unit commitment decision is used to learn the sample set offline, the "input-output" mapping relationship between the load matrix and the low-dimensional unit start-stop state encoding matrix can be obtained. Subsequently, the No.91-No.93 load matrices of the sample set can be decided online.

[0196] The online decision of method 2 is combined with the unit commitment decision method based on Benders decomposition to solve the unit day-ahead schedule of the No.91-No.93 load matrix as follows: Figure 4 shown.

[0197] Depend on Figure 4 It can be seen that in the test set composed of No. 91-No. 93 load matrices, the unit day-ahead power generation schedules solved by the method of the present invention and the unit combination decision method based on Benders decomposition are consistent. This shows that the unit combination decision method based on high-dimensional sample XOR sparse coding strategy proposed in the present invention can fully learn the mapping relationship between the load matrix and the unit start-stop state variable matrix from the unit combination decision sample set, and realize that the encoder-decoder network model can formulate correct unit start-stop plans when facing different load matrix scenarios.

[0198] In order to further verify the correctness of the present invention, the No. 91-No. 93 load matrix of the test set and the unit start-stop state variable matrix solved by method 2 are input into the OPF model to solve the active power of the unit. The operating cost comparison of the two methods is shown in Table 3.

[0199] Table 3 Comparison of unit operating costs of the method of the present invention and the Benders decomposition method

[0200]

[0201] It can be observed that the unit active power operating cost obtained by method 2 is the same as the unit combination method based on Benders decomposition. The reason is that for the same test sample, such as the No. 91 load matrix, since the two methods use the same OPF model to solve the unit active power, the input load matrix and the high-dimensional unit start-stop state variable matrix are also the same. The differences in the No. 91-No. 93 load matrices also lead to slightly different calculated active powers and different system operating costs.

[0202] (2) Effectiveness analysis of XOR sparse coding method for high-dimensional samples

[0203] In order to analyze the influence of the XOR sparse coding method for high-dimensional samples of the present invention on the training efficiency of the encoder-decoder network model of unit combination decision, Table 4 gives the offline learning time of Method 1 and Method 2 on the No.91-No.93 load matrix training set and the decision accuracy on the No.91-No.93 load matrix test set.

[0204] Table 4 Performance comparison of XOR sparse coding methods for high-dimensional samples

[0205]

[0206] As shown in Table 4, the unit start and stop of the two methods are consistent, with an accuracy of 100%. However, the offline learning time of method 2 in the training set is 351s less than that of method 1, and the XOR sparse encoding time of its high-dimensional unit start and stop state variable matrix is ​​only 0.1266s. This result clearly shows that the proposed sparse XOR encoding method for high-dimensional samples can effectively compress the high-dimensional unit start and stop state variable matrix from T×N dimensions to T×1 dimensions. Although the matrix encoding process consumes a certain amount of time, it can significantly reduce the number of variables that need to be calculated by the encoder-decoder network model during offline learning.

Claims

1. A unit commitment decision method based on high-dimensional sample XOR sparse coding, characterized in that: The following steps are involved: Step 1: Perform coding preprocessing on the high-dimensional unit start-stop state variable matrix to obtain a low-dimensional unit start-stop state coding matrix; Step 2: Pair the load matrix with the high-dimensional unit start-stop state variable matrix to form a high-dimensional unit combination decision sample set of the invention, perform XOR sparse coding on the high-dimensional unit start-stop state variable matrix to reduce the dimension, and pair the load matrix with the low-dimensional unit start-stop state coding matrix to form a low-dimensional unit combination decision sample set of the invention; Step 3: Construct an encoder-decoder network model suitable for unit commitment decision-making, and use the low-dimensional unit commitment decision sample set obtained in step 2 to train and solve the encoder-decoder network model of unit commitment decision-making; The above steps are used to realize unit combination decision.

2. The method according to claim 1, characterized in that In step 1, the coding preprocessing of the high-dimensional unit start-stop state variable matrix includes the following sub-steps: Step 1-1): Identify the k-th high-dimensional unit start-stop state variable matrix The dimensions, number of rows T and number of columns N; Where: is the k-th high-dimensional unit start-stop state variable matrix, where the subscript 0 represents the state that is not XOR-sparse and not encoded; is the start / stop state 0-1 variable of unit n at time t in the matrix. The value range of time t is t∈[1,T], and the value range of unit n is n∈[1,N]; the number of rows T is the time of unit combination decision; the number of columns N is the number of generating units; Step 1-2): Take the i-th row in the k-th high-dimensional unit start-stop state variable matrix, i∈[1,T], vector As the reference row vector, perform XOR operation on each element of the remaining row vector bit by bit with the corresponding element of the i-th row; Where: is the reference row vector composed of the start-stop state variables of N units in the k-th high-dimensional unit start-stop state variable matrix; To form the base row vector The element represents the start / stop state of unit n at time i, a 0-1 variable; Taking the remaining rows, row t, as an example, then is the t-th row vector With the reference row vector The row vector after bit-by-bit XOR operation; is the exclusive OR operator; Step 1-3): Compare all remaining row vectors with the reference row vector After performing XOR operation bit by bit, the elements contained in all row vectors can be combined into a new high-dimensional unit start-stop state variable matrix The subscript i indicates the state where the i-th row vector is XOR-sparse; Where: is a new high-dimensional unit start-stop state variable matrix obtained by performing XOR sparse operation bit by bit with the i-th row vector as the reference row; The row vector after performing bit-by-bit XOR sparse operations Assign a decimal value to encode, assuming for but: Step 1-4): When the k-th high-dimensional unit start-stop state variable matrix The base row vector The row number i of the sample is completed once from 1 to T, and the XOR sparse coding cycle iteration is regarded as the high-dimensional unit start-stop state variable matrix completing a complete XOR sparse coding process, and the XOR sparse coding result Z of the sample is obtained. k ; Where: is the kth high-dimensional unit start-stop state variable matrix After a complete XOR sparse coding cycle, the k-th low-dimensional unit start-stop state coding matrix Z is obtained k ; Step 1-5): If the XOR sparse coding results of two consecutive high-dimensional samples are the same, it is considered to have reached the end state and proceed to step 1-6); otherwise, go to step 1-2); Step 1-6): Output the low-dimensional unit start-stop state encoding matrix Z k ; Where: z i The result is decimal encoding.

3. The method according to claim 2, characterized in that In step 2, a low-dimensional unit commitment decision sample set S is constructed low , the sample data includes: load matrix P and pre-processed low-dimensional unit start-stop state encoding matrix Z, including the following sub-steps: Step 2-1): Based on the load data and unit dispatching plan collected by the monitoring equipment during the actual operation of the power grid, the load matrix P and the high-dimensional unit start-stop state variable matrix U are paired to form the high-dimensional unit combination decision sample set S of the invention. high , usually expressed as: Where: is the high-dimensional unit commitment decision sample set S high The kth load matrix of k is the high-dimensional unit commitment decision sample set S high The k-th high-dimensional unit start-stop state variable matrix, where U k The dimension is T×N; K is the total number of unit commitment decision samples in the sample set; Step 2-2): Use the method in step 1 to reduce the dimension of the high-dimensional unit start-stop state variable matrix U by XOR sparse coding, and pair the load matrix P with the low-dimensional unit start-stop state coding matrix Z to form the invented low-dimensional unit combination decision sample set S low , usually expressed as: Where: Z k is the low-dimensional unit commitment decision sample set S high The k-th low-dimensional unit start-stop state encoding matrix, where Z k The dimension is T×1.

4. The method according to claim 3, characterized in that In step 3, an encoder-decoder network model suitable for unit commitment decision-making is constructed, and the low-dimensional unit commitment decision sample set S in step 2 is used low Training and solving the encoder-decoder network model for unit commitment decision-making includes the following sub-steps: Step 3-1), construct an encoder-decoder network model suitable for unit combination decision-making, the input of the network model is the load matrix P of dimension T×1, and the output is the low-dimensional unit start-stop state encoding matrix Z of dimension T×1; wherein, the encoder network model of unit combination decision-making converts the load matrix P into a fixed-length vector expression C, and the decoder network model of unit combination decision-making converts the fixed-length vector expression C into a low-dimensional unit start-stop state encoding matrix Z, realizing the process from encoding to decoding of the unit combination sample matrix, and the main sub-steps are as follows: Step 3-1-1): The encoder network model of unit combination decision reads the load matrix P according to the time step; at time t, the input features of the GRU neurons of the encoder network model include: the current input load P t and the hidden layer state h of the previous GRU neuron at time t-1 t-1 ; The hidden layer state h of the GRU neuron at time t can be obtained t , which is encoded into the intermediate state C by the encoder network model of the unit combination decision t Then input the next GRU neuron, specifically: h t =f(h t-1 ,P t ) (10) C t =h t (11) Where: h t is the hidden layer state of the GRU neuron in the encoder network model for unit combination decision at time t; h t-1 is the hidden layer state of the previous GRU neuron at time t-1; P t is the load information at time t; f is the transformation function of the hidden layer of the recurrent neural network; C t is the intermediate state after being encoded by the encoder network model at time t; Step 3-1-2): The intermediate state of the encoder network model of the unit commitment decision at time T is the concatenation of the hidden layer states of all time steps: C T =q({h1,h2,…,h T }) (12) Where: C T is the coded information output at time T; q is a custom nonlinear function; Step 3-1-3): The decoder network model of unit commitment decision outputs the low-dimensional unit start-stop state encoding matrix Z according to the time step; the low-dimensional unit start-stop state encoding matrix Z at time t-1 is converted into t-1 and the hidden layer state s′ of the previous GRU neuron t-1 And the intermediate state C after being decoded by the decoder network model at time t t As the input of the GRU neuron of the decoder network model at time t, the hidden layer state s′ of the GRU neuron of the decoder network model at time t is output t-1 ; s′ t =g(s′ t-1 ,Z t-1 ,C t ) (13) C t =s′ t (14) Where: s′ t The hidden layer state of the GRU neuron in the decoder network model for unit combination decision-making at time t; Z t-1 is the low-dimensional unit start-stop state encoding matrix output at time t-1; s′ t-1 is the hidden layer state of the previous GRU neuron at time t-1; C t is the intermediate state after being decoded by the decoder network model at time t; g is the transformation function of the hidden layer of the recurrent neural network; Step 3-1-4): The probability information decoded and output at time T by the decoder network model of unit commitment decision is based on the set of low-dimensional unit start-stop state encoding matrices {Z1, Z2, …, Z T-1 } and the intermediate state C after being decoded by the decoder network model at the previous time step T-1 Joint decision: Where: p(Z T ) is the probability that the decoder outputs the decoded information at the last time step T; {Z1, Z2, …, Z T-1 The set C is the set of low-dimensional unit start-stop state encoding matrices output at the previous T-1 time; T-1 is the intermediate state after being decoded by the decoder at the previous time step; Step 3-2), using the low-dimensional unit commitment decision sample set S in step 2 low The encoder-decoder network model suitable for unit commitment decision established in training step 3-1) can obtain the general mapping relationship of unit commitment decision: (Z)=p(P L ) (16) Where: P L is the load matrix; Z is the low-dimensional unit start-stop state encoding matrix; p(*) is the load matrix P L The probability mapping model with the low-dimensional unit start-stop state encoding matrix Z; The trained encoder-decoder network model for unit commitment decision-making can learn the mapping between the daily load matrix and the unit start-stop state variable encoding matrix in an offline state; Step 3-3), input the daily load matrix and the unit start-stop state variable matrix to the optimal power flow OPF model, and solve the output power of the generator based on the optimal power flow OPF model, including the following sub-steps: Step 3-3-1): The objective function of the OPF model is to minimize the operating cost of the unit: Where: F(P G ) is the operating cost of a conventional generator set; a i ,b i ,c i are the coal consumption coefficients of conventional power generating units; P Git is the output power of conventional generator set i at time t, U Git is the start and stop status of conventional generator set i at time t; Step 3-3-2): Decision variable P Git Power balance constraints, output upper and lower limit constraints, climbing / slip rate constraints, and network security constraints must also be met to ensure the safe operation of the system: Where: P Lt is the daily load at time t; and are the lower and upper limits of active output of unit i respectively; and are the sliding rate and climbing rate of unit i respectively; P lmax is the maximum transmission capacity allowed by line l in the base state; G l-i is the generator output power transfer distribution factor of unit i to line l; B is the number of nodes in the power grid; D bt is the load of node b in the power grid during period t; Step 3-3-3): Output the active power of the unit according to the OPF model.

5. An encoder-decoder network model for unit commitment decision-making, characterized in that: It includes an encoder network model for unit combination decision and a decoder network model for unit combination decision; the input of the encoder-decoder network model for unit combination decision is a load matrix P with a dimension of T×1, and the output is a low-dimensional unit start-stop state encoding matrix Z with a dimension of T×1; wherein, the encoder network model for unit combination decision converts the load matrix P into a fixed-length vector expression C, and the decoder network model for unit combination decision converts the fixed-length vector expression C into a low-dimensional unit start-stop state encoding matrix Z, so as to realize the process from encoding to decoding of the unit combination sample matrix.

6. The method according to claim 5, characterized in that The encoder-decoder network model for unit commitment decision-making, when working, includes the following steps: Step S1): The encoder network model for unit commitment decision reads the load matrix P according to the time step; Specifically, at time t, the input features of the GRU neurons in the encoder network model include: the current input load P t and the hidden layer state h of the previous GRU neuron at time t-1 t-1 ; The hidden layer state h of the GRU neuron at time t can be obtained t , which is encoded into the intermediate state C by the encoder network model of the unit combination decision t Then input the next GRU neuron, which is specifically expressed as: h t =f(h t-1 ,P t ) (10) C t =h t (11) Where: h t is the hidden layer state of the GRU neuron in the encoder network model for unit combination decision at time t; h t-1 is the hidden layer state of the previous GRU neuron at time t-1; P t is the load information at time t; f is the transformation function of the hidden layer of the recurrent neural network; C t is the intermediate state after being encoded by the encoder network model at time t; Step 2): The encoder network model of the unit commitment decision encodes and outputs the intermediate state at time T, and the intermediate state is the concatenation of the hidden layer states of all time steps: C T =q({h1,h2,…,h T }) (12) Where: C T is the coded information output at time T; q is a custom nonlinear function; Step 3): The decoder network model of unit commitment decision outputs a low-dimensional unit start-stop state encoding matrix Z according to the time step; Specifically: the low-dimensional unit start-stop state encoding matrix Z at time t-1 t-1 and the hidden layer state s′ of the previous GRU neuron t-1 And the intermediate state C after being decoded by the decoder network model at time t t As the input of the GRU neuron of the decoder network model at time t, the hidden layer state s′ of the GRU neuron of the decoder network model at time t is output t-1 ; s′ t =g(s′ t-1 ,Z t-1 ,C t ) (13) C t =s′ t (14) Where: s′ t The hidden layer state of the GRU neuron in the decoder network model for unit combination decision-making at time t; Z t-1 is the low-dimensional unit start-stop state encoding matrix output at time t-1; s′ t-1 is the hidden layer state of the previous GRU neuron at time t-1; C t is the intermediate state after being decoded by the decoder network model at time t; g is the transformation function of the hidden layer of the recurrent neural network; Step 4): The decoder network model of the unit commitment decision decodes and outputs probability information at time T, and the probability information is based on the low-dimensional unit start-stop state encoding matrix set {Z1, Z2, ..., Z T-1 } and the intermediate state C after being decoded by the decoder network model at the previous time step T-1 Joint decision: Where: p(Z T ) is the probability that the decoder outputs the decoded information at the last time step T; {Z1, Z2, …, Z T-1 The set C is the set of low-dimensional unit start-stop state encoding matrices output at the previous T-1 time; T-1 It is the intermediate state after being decoded by the decoder at the previous time step.

7. The method according to claim 6, characterized in that The trained encoder-decoder network model for unit combination decision-making can learn the mapping between the daily load matrix and the unit start-stop state variable encoding matrix in an offline state, so that other models can obtain the unit output power by inputting the daily load matrix and the unit start-stop state variable matrix.

Images