A deep learning-based power system economic dispatch network loss calculation method

Abstract

The present application relates to a kind of based on deep learning's power system economic dispatch network loss calculation method, belong to power system economic dispatch technical field, including the following steps: step 1, using differential evolution algorithm simulates and generates the training data of network loss calculation model;Step 2, build multi-head multilayer perceptron network loss calculation model and carry out model training;Step 3, call multi-head multilayer perceptron network loss calculation model and carry out power system economic dispatch calculation.The present application replaces the traditional calculation method by the method of deep learning, greatly improves the calculation speed of network loss, and the multilayer perceptron model is improved to form multi-head multilayer perceptron model, realizes network loss calculation under different load demand without retraining;Input is unit output data and each node load demand, so that the network loss calculation method is more suitable for power system economic dispatch.

Description

TECHNICAL FIELD

[0001] The application relates to a deep learning-based power system economic dispatch network loss calculation method and belongs to the technical field of power system economic dispatch. BACKGROUND

[0002] Power system economic dispatch is a way of reasonably utilizing energy and equipment. In power system economic dispatch, network loss accounts for a large proportion, and network loss constraints need to be considered in power balance. Traditional network loss calculation methods include alternating current flow calculation, b coefficient method and network loss sensitivity method, but due to the limitations of solving methods and hardware devices, it is difficult to meet the requirements of accuracy or speed. Therefore, the application provides a network loss calculation method suitable for power system economic dispatch. SUMMARY

[0003] The application aims to provide a deep learning-based power system economic dispatch network loss calculation method, improve the multilayer perceptron model, form a multi-head multilayer perceptron model, and realize network loss calculation under different load demands without retraining; and input unit output data and node load data to improve the calculation speed.

[0004] In order to achieve the above-mentioned purpose, the technical scheme adopted by the application is:

[0005] A deep learning-based power system economic dispatch network loss calculation method comprises the following steps:

[0006] Step 1: adopt a differential evolution algorithm to simulate and generate training data of a network loss calculation model;

[0007] Step 2: build a multi-head multilayer perceptron network loss calculation model and perform model training;

[0008] Step 3: call the multi-head multilayer perceptron network loss calculation model to perform power system economic dispatch calculation.

[0009] The further improvement of the technical scheme of the application is that the specific process of step 1 is:

[0010] In order to ensure that the network loss calculation method can be applied to different load requirements, the differential evolution algorithm is used to simulate the generator output data under different load requirements; in the operation process of the differential evolution algorithm, the unit output data generated in each iteration process is saved, and the corresponding network loss is calculated by calling the alternating current flow calculation; in order to meet the diversity of unit output data, the cross factor CR and the mutation factor F are set to larger parameters to increase the diversity of the population; at the same time, the result of differential evolution must be the optimal solution; in order to make the multi-head multi-layer perceptron model applicable to different load requirements, the load data is generated with a gradient interval of 0.05 times from 0.4 times of the maximum total output of the unit to 0.85 times of the maximum total output of the unit to generate training data under different loads.

[0011] The further improvement of the technical scheme of the application is that the specific process of step 2 is:

[0012] The multi-layer perceptron understands the mapping relationship inside the data through historical data, and obtains powerful computing power through the cooperation between neurons; the core of the multi-layer perceptron is the neuron, which receives external input data or the output of other neurons and gives it a weight w; then a bias value b is added to the bias amount, and the output Y is calculated through the activation function f, which is called the forward propagation process:

[0013] Y i =f i (w i Y i-1 +b i ) (8)

[0014] Among them, Y i is the output of the i-th layer of the multi-layer perceptron, w i is the weight vector of the i-th layer, and b i is the bias vector of the i-th layer:

[0015] Without the activation function, the information transmission between neurons becomes a linear relationship, and the neurons lose activity, so it is essential to choose a suitable activation function; the linear rectifier activation function is used in the hidden layer; the linear rectifier activation function does not have a saturation zone, and there is no problem of gradient disappearance; the linear activation function is used in the output layer, and the batch gradient descent method is used to train the multi-layer perceptron network; the purpose of the batch gradient descent method is to minimize the loss function value, that is, the calculated value is closer to the true value, and this process is called the back propagation process:

[0016]

[0017] Among them, ( is w i and b iThe function is the updated neuron output vector, and alpha is the gradient descent step size.

[0018] Since the unit output data is generated by proportional gradient simulation, the traditional multi-layer perceptron model will have a slight deviation from the true value when calculating the network loss data between two proportional gradients; in order to minimize the error as much as possible, a multi-head multi-layer perceptron model is proposed to reduce the error; the multi-head multi-layer perceptron is composed of multiple parallel units of multi-layer perceptron, different units are independently trained, and the output of the model is obtained by weighted average of the outputs of multiple multi-layer perceptrons;

[0019] In order to ensure the independence between different multi-layer perceptron units, the training data is randomly divided, and the first 70% of data of each group of data is saved for training different multi-layer perceptron units; under different load requirements, the network loss changes greatly; in order to avoid the phenomenon that large data covers small data, the weight coefficient determination method with the minimum relative error is used to determine the weight of different calculation heads.

[0020] The further improvement of the technical scheme of the application is that the specific process of step 3 is:

[0021] In the economic dispatching of the power system, the objective function of the economic dispatching problem takes the minimum coal consumption of the thermal power unit as the objective function:

[0022]

[0023] Wherein, F is the total power generation cost of economic dispatching, a i , b i , c i are the cost coefficients of the power generating unit i, P G is the output power of the unit i, and N is the total number of units;

[0024] After considering the valve point effect of the system, the objective function is as follows:

[0025]

[0026] Wherein, e i and f i are the cost coefficients of the i-th power generator reflecting the valve point loading effect, P i,min is the minimum output power of the i-th unit;

[0027] The constraint condition considers the maximum and minimum output of the unit and the load power balance;

[0028]

[0029] Wherein, P GD is the total load demand, and P G,LOSS is the total line loss.

[0030] P i,min ≤P G,i ≤P i,max (15)

[0031] where, p i,min is the minimum value of the i th unit output, p i，max is the maximum value of the i th unit output;

[0032] Due to the physical characteristics of the thermal power generating unit, in order to avoid the bearing vibration amplification of the thermal power generating unit in some operation interval, the power generation prohibited interval constraint and the system climbing constraint are considered;

[0033] The differential evolution algorithm is used to solve the economic dispatching problem of the power system, the differential evolution process is divided into population initialization, variation, crossover and selection, the multi-head multilayer perceptron network loss calculation model is called after the crossover link to calculate the network loss, and the optimal selection of unit output is carried out under the condition of considering the network loss.

[0034] The further improvement of the technical scheme of the application is that the weight of different calculation heads is determined by using the weight coefficient determination method of the minimum relative error, that is:

[0035]

[0036]

[0037] where, y i is true value, m is the data amount used for calculating weight, f(i) is the i th calculation value of the t th head, and ω t is the weight value of the t th head.

[0038] The further improvement of the technical scheme of the application is that the power generation prohibited interval constraint is:

[0039]

[0040] wherein, is the lower limit of the first prohibited output interval of the i th unit, is the upper limit of the k th prohibited output interval of the i th unit, and poz i is the total number of prohibited output intervals of the i th unit.

[0041] The further improvement of the technical scheme of the application is that the system climbing constraint is:

[0042] P i,t -P i,t-1 ≤RU g (17)

[0043] P i,t-1-P i,t ≤RD g (18)

[0044] wherein, P i,t and P i,t-1 are the power generation of the unit i at t and t-1 time, RU g and RD g are the output of the thermal power unit per unit time.

[0045] Due to the adoption of the above technical scheme, the technical effects obtained by the present application are:

[0046] The present application replaces the traditional calculation method with the method of deep learning, greatly improves the calculation speed of network loss, and improves the multi-layer perceptron model to form a multi-head multi-layer perceptron model, so as to realize the network loss calculation under different load demands without retraining.

[0047] The input of the present application is the unit output data and the load demand of each node, so that the network loss calculation method is more suitable for economic dispatching of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is the overall framework diagram of the present application;

[0049] Figure 2 is the training process diagram of the neural network in the present application;

[0050] Figure 3 is the economic dispatching flow chart. DETAILED DESCRIPTION

[0051] The present application will be further described in detail below in combination with the drawings and specific embodiments:

[0052] In the economic dispatching of the power system, the objective function of the economic dispatching problem is generally to minimize the coal consumption of the thermal power unit as the objective function.

[0053]

[0054] wherein, F is the total power generation cost of economic dispatching, a i , b i , c i are the cost coefficients of the power generation unit i, P G is the output power of the unit i, and N is the total number of units.

[0055] After considering the valve point effect of the system, the objective function becomes the following form:

[0056]

[0057] wherein, ei and f i To reflect the cost coefficient of the i-th generator that reflects the valve point loading effect, P i,min Let be the minimum output power of the i-th unit.

[0058] The constraints take into account the unit's maximum and minimum output and load power balance.

[0059]

[0060] Among them, P GD For total load demand, P G,LOSS This represents the total line loss.

[0061] P i,min ≤P G,i ≤P i,max (4)

[0062] Where, p i,min p is the minimum output value of the i-th unit. i，max It is the maximum value output by the i-th unit.

[0063] Due to the physical characteristics of thermal power generating units, in order to avoid amplification of bearing vibration in certain operating ranges, a power generation prohibition range constraint was considered.

[0064]

[0065] in, This is the lower limit of the first prohibited output range for the i-th unit. poz is the upper limit of the k-th prohibited output range of the i-th unit. i This represents the total number of output-prohibited intervals for the i-th generator unit.

[0066] Consider the system's ramping constraints:

[0067] P i,t -P i,t-1 ≤RU g 6

[0068] P i,t-1 -P i,t ≤RD g 7

[0069] Among them, P i,t and P i,t-1 Let RU represent the power generation of unit i at times t and t-1, respectively. g and RD g These represent the increase or decrease in the output of thermal power units per unit time.

[0070] A deep learning-based method for calculating network losses in economic dispatch of power systems, such as...Figure 1 As shown, comprising the following steps:

[0071] Step 1, using differential evolution algorithm to simulate the training data of network loss calculation model

[0072] According to the maximum value of the total output of the system unit, the load demand is set with 0.5 times gradient, and the training data of the network loss calculation model is generated, so as to ensure that the network loss calculation model can realize the network loss calculation under different load demands.

[0073] Due to the lack of historical data of power system, under different load demands, differential evolution algorithm is used to simulate the training data of 30-node improved system economic dispatch network loss calculation model. After the cross-link of differential evolution algorithm, the unit output data is obtained, and the network loss at this time is accurately calculated by calling the alternating current flow algorithm. Repeat the above steps in each iteration process of differential evolution algorithm, save the unit output and the corresponding network loss at each time to generate training data. Randomly shuffle the training data and divide it into 4 groups for training of different calculation heads to ensure the independence between heads.

[0074] Step 2, build multi-head multi-layer perceptron network loss calculation model and train the model

[0075] Multi-layer perceptron can understand the mapping relationship inside the data through historical data, and obtain powerful computing ability through the cooperation between neurons. The core of multi-layer perceptron is neuron, which receives external input data or the output of other neurons and gives it a weight w; Then add the bias value b of the bias, calculate the output Y through the activation function f, this process is called the forward propagation process.

[0076] Y i =f i (w i Y i-1 +b i ) (8)

[0077] Where, Y i is the output of the i-th layer of multi-layer perceptron, w i is the weight vector of the i-th layer, and b i is the bias vector of the i-th layer.

[0078] Without activation function, the information transmission between neurons becomes a linear relationship, and the neurons lose activity, so it is essential to choose a suitable activation function. The hidden layer adopts a linear rectifier activation function (ReLu). The ReLu activation function has no saturation zone, and there is no problem of gradient disappearance. Since the invention deals with a regression problem, the output layer adopts a linear activation function (Linear). The batch gradient descent method is used to train the multilayer perceptron network. The purpose of the batch gradient descent method is to minimize the loss function value, that is, the calculated value is closer to the true value, and this process is called the back propagation process:

[0079]

[0080] wherein, ( is the updated neuron output vector, and a is the gradient descent step size. i and b i are functions of w i .

[0081] Since the unit output data is generated by proportional gradient simulation, the traditional multilayer perceptron model will produce a slight deviation from the true value when calculating the network loss data between two proportional gradients. In order to minimize the error as much as possible, a multi-head multilayer perceptron model is proposed to reduce the error. The multi-head multilayer perceptron is composed of multiple parallel units of multilayer perceptron, and different units are trained independently. The output of the model is obtained by weighted average of the outputs of multiple multilayer perceptrons.

[0082] In order to ensure the independence between different multilayer perceptron units, the training data is randomly divided, and the first 70% of data of each group of data is saved for training different multilayer perceptron units. Under different load requirements, the network loss changes greatly. In order to avoid the phenomenon that large data covers small data, the weight coefficient determination method with the minimum relative error is used to determine the weight of different calculation heads.

[0083]

[0084]

[0085] wherein, y i is the true value, m is the amount of data used for calculating the weight, f(i) is the i th calculation value of the t th head, and ω t is the weight of the t th head.

[0086] Considering the complexity and accuracy of the model calculation, the number of heads of the multi-head multilayer perceptron network loss calculation model is determined to be 4, the hidden layer of each independent calculation head is set to 3 layers, and the number of neurons is 64, 32 and 8 respectively. Each calculation head is independently trained 300 times using the batch gradient descent method, and the optimal model result is saved.

[0087] The weight parameters of each independent calculation head are calculated by using the weight coefficient determination method with the minimum relative error, and the data between 900 different load demand gradients are randomly selected for weight calculation. The training process of the neural network is referenced Figure 2 .

[0088] Step 3, calling the multi-head multi-layer perceptron loss calculation model to perform economic dispatch calculation of the power system

[0089] In the economic dispatch of the power system, the objective function of the economic dispatch problem is generally to minimize the coal consumption of the thermal power unit as the objective function.

[0090]

[0091] Where F is the total generation cost of economic dispatch, a i , b i , c i are the cost coefficients of the i-th generator, P G is the output power of the i-th unit, and N is the total number of units.

[0092] After considering the valve point effect of the system, the objective function becomes the following form:

[0093]

[0094] Where e i and f i are the cost coefficients of the i-th generator reflecting the valve point loading effect, P i,min is the minimum output power of the i-th unit.

[0095] The constraint condition considers the maximum and minimum output of the unit and the load power balance.

[0096]

[0097] Where P GD is the total load demand, and P G,LOSS is the total line loss.

[0098] P i,min ≤ P G,i ≤ P i,max (15)

[0099] Where p i,min is the minimum value of the i-th unit output, and p i，max is the maximum value of the i-th unit output.

[0100] Due to the physical characteristics of the thermal power unit itself, in order to avoid the amplification of bearing vibration of the thermal power unit in some operating range, the generation prohibited interval constraint is considered.

[0101]

[0102] wherein, is the lower limit of the first prohibited output interval of the i-th unit, is the upper limit of the k-th prohibited output interval of the i-th unit, poz i is the total number of prohibited output intervals of the i-th unit.

[0103] Considering the ramping constraints of the system:

[0104] P i,t -P i,t-1 ≤RU g (17)

[0105] P i,t-1 -P i,t ≤RD g (18)

[0106] wherein P i,t and P i,t-1 are the power generation of the i-th unit at t and t-1, RU g and RD g are the power output increase and decrease of the thermal power unit per unit time.

[0107] The differential evolution algorithm is used to solve the economic dispatching problem of the power system, and the differential evolution process is divided into population initialization, mutation, crossover and selection. After the crossover link, the network loss calculation model is called to calculate the network loss. The optimal selection of unit output is carried out under the consideration of network loss.

[0108] Four different load demands are randomly selected to evaluate the network loss calculation model. The accuracy of the network loss calculation model is evaluated under the load demands of 160 MW, 189.2 MW, 240 MW and 275 MW. Under different load demands, 1200 sets of unit output data are generated by using the differential evolution method, and the network loss under different unit output conditions is calculated by using the multi-head multi-layer perceptron network loss calculation model. The calculation results are compared with the alternating current flow calculation results. Since most of the unit output values in the late iteration of the differential evolution algorithm satisfy the equality constraints of economic dispatching, the last 300 sets of data are selected for model evaluation. The r 2 values and calculation errors of the last 300 sets of data are shown in the following table.

[0109] Table 1 Model calculation accuracy evaluation results

[0110]

[0111] r 2The values are all greater than 0.997, indicating that the fitting degree of the calculated curve to the actual curve is very high. From the average error value and the maximum error value of the network loss under the load demands of 160 MW, 189.2 MW, 240 MW and 275 MW, it can be seen that the calculated value of the network loss is highly similar to the true value. With the convergence of the differential evolution algorithm, the network loss error corresponding to the unit output value meeting the requirements of economic dispatch of the power system is very small. At the end of the differential evolution iteration, although the change of the unit output is very small, the multi-head multi-layer perceptron model can still perceive the very small change and perform accurate network loss calculation. Using the multi-head multi-layer perceptron, the average calculation speed of the multi-layer perceptron network loss calculation model is 5.1 times faster than the alternating current flow in 1000 network loss calculations of the 30-node improved system model. It can be seen that the more the number of iterations in the economic dispatch process, the greater the advantage of the algorithm.

[0112] Table 2 network loss calculation time

[0113]

[0114] The differential evolution algorithm is used to calculate the economic dispatch of the 30-node improved system, and the Newton alternating current flow algorithm and the network loss calculation method in the present application are used to calculate the network loss when considering the network loss constraint condition. The economic dispatch process is shown in Figure 3 , and the economic dispatch results are shown in the following table.

[0115] Table 3 economic dispatch results

[0116]

[0117]

[0118] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the foregoing embodiments of the present application have been described in detail, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A deep learning-based method for calculating network losses in economic dispatch of a power system, characterized in that, Includes the following steps: Step 1: Use the differential evolution algorithm to simulate and generate training data for the network loss calculation model; The specific process of step 1 is as follows: Due to the lack of historical data, to ensure that the network loss calculation method is applicable to different load demands, a differential evolution algorithm is used to simulate generator output data under different load demands. During the operation of the differential evolution algorithm, the generator output data generated in each iteration is saved, and the corresponding network loss is calculated by calling the AC power flow calculation. In order to meet the diversity of generator output data, the crossover factor CR and the mutation factor F are set to larger parameters to increase the diversity of the population. At the same time, the result of differential evolution must be the optimal solution. In order to make the multi-head multilayer perceptron model applicable to different load requirements, the load data is generated with a starting point of 0.4 times the maximum total generator output and an ending point of 0.85 times the maximum total generator output, with a gradient interval of 0.05 times to generate training data under different loads. Step 2: Build a multi-head, multi-layer sensor network loss calculation model and train the model; The specific process of step 2 is as follows: Multilayer perceptrons (MLPs) learn about the mapping relationships within data through historical data and gain powerful computing capabilities through the collaboration between neurons. The core of the MLP is the neuron, which receives external input data or the output of other neurons and assigns them weights w. Then, a bias term with a bias value of b is added, and the output Y is calculated through the activation function f. This process is called the preceding term propagation process. ； in, This is the output of the i-th layer of the multilayer perceptron. Let i be the weight vector of the i-th layer. Here is the bias vector for the i-th layer: Without an activation function, information transfer between neurons becomes linear, and neurons lose their activity. Therefore, choosing a suitable activation function is crucial. The hidden layers use a linear rectified activation function. Linear rectified activation functions do not have a saturation region and do not suffer from the gradient vanishing problem. The output layer uses a linear activation function, and the multilayer perceptron network is trained using batch gradient descent. The purpose of batch gradient descent is to minimize the loss function value, i.e., to make the calculated value closer to the true value. This process is called backpropagation. ； in, yes and The function, Here is the updated neuron output vector, and α is the gradient descent step size; Since the unit output data is generated by proportional gradient simulation, the traditional multilayer perceptron model will produce a slight deviation from the true value when calculating the network loss data between two proportional gradients. In order to minimize the error, a multi-head multilayer perceptron model is proposed. The multi-head multilayer perceptron is composed of parallel units of multiple multilayer perceptrons. Different units are trained independently, and the output of the model is obtained by weighted average of the outputs of multiple multilayer perceptrons. To ensure the independence between different multilayer perceptron units, the training data is randomly divided, and the first 70% of the data in each group is saved for training different multilayer perceptron units. The network loss varies greatly under different load requirements. To avoid the phenomenon of large data covering small data, the weight coefficient determination method with the smallest relative error is used to determine the weights of different computation heads. Step 3: Call the multi-head, multi-layer sensor network loss calculation model to perform power system economic dispatch calculation; The specific process of step 3 is as follows: In the economic dispatch of power systems, the objective function of the economic dispatch problem is to minimize the coal consumption of thermal power units. ； Where F represents the total power generation cost under economic dispatch. , , Let i be the cost coefficient of generator set i. Where N is the output power of unit i, and N is the total number of units; After considering the system's threshold effect, the objective function takes the following form: ； in, and To reflect the cost coefficient of the i-th generator that reflects the valve point loading effect, Let be the minimum output power of the i-th unit; The constraints take into account the unit's maximum and minimum output and load power balance; ； in, For total load demand, Total line loss; ； in, It is the minimum output value of the i-th unit. It is the maximum value output by the i-th unit; Due to the physical characteristics of thermal power generating units, in order to avoid amplification of bearing vibration in certain operating ranges, constraints on prohibited power generation ranges and system ramping constraints were considered. The differential evolution algorithm is used to solve the economic dispatch problem of the power system. The differential evolution process is divided into population initialization, mutation, crossover and selection. After the crossover stage, the network loss calculation model of multi-head multi-layer sensor is called to calculate the network loss. The optimal selection of unit output is made considering the network loss.

2. The method for calculating network losses in power system economic dispatch based on deep learning according to claim 1, characterized in that: The weights of different computation heads are determined using a weighting coefficient determination method that minimizes relative error: ； ； in, Here, m represents the true value, and m represents the amount of data used to calculate the weights. For the i-th calculated value of the t-th head, Let t be the weight of the t-th head.

3. The method for calculating network losses in power system economic dispatch based on deep learning according to claim 1, characterized in that: The constraint for the power generation prohibition interval is: ； in, This is the lower limit of the first prohibited output range for the i-th unit. This represents the upper limit of the k-th prohibited output interval for the i-th unit. This represents the total number of output-prohibited intervals for the i-th generator unit.

4. The method for calculating network losses in power system economic dispatch based on deep learning according to claim 1, characterized in that: The system's ramp constraint is: (17) (18) in and Let be the power generation of unit i at times t and t-1, respectively. and These represent the increase or decrease in the output of thermal power units per unit time.

Images