Hydrothermal power generation system scheduling optimization method based on transfer learning and GAN

Abstract

The invention discloses a hydrothermal power generation system scheduling optimization method based on transfer learning and a GAN, and the method comprises the steps: combining a randomly generated scheduling scheme population with an optimization scheduling scheme population of a previous power demand when a current power demand is a second power demand; obtaining an optimal scheduling scheme population of the current power demand on the basis of a non-dominated sorting genetic algorithm to perform hydrothermal power generation system scheduling, and when the current power demand is a third or more power demand, extracting experience of a historical optimal scheduling scheme population on the basis of a transfer learning thought; generating a new scheduling scheme population with historical experience in combination with a GAN (Generative Adversarial Network) and standard normal distribution, and obtaining an optimized scheduling scheme population of the current power demand through fusion and a non-dominated sorting genetic algorithm to schedule the hydrothermal power generation system; the invention has the advantage of high stability in balancing the total fuel cost of the minimum thermal power generator set and the emission load of the minimum thermal power generator set for the overall power demand.

Description

Technical Field

[0001] The present invention relates to a scheduling optimization method for a hydrothermal power generation system, and in particular to a scheduling optimization method for a hydrothermal power generation system based on transfer learning and GAN. Background Art

[0002] In a hydrothermal power generation system, a hydro network composed of hydroelectric generating units and a thermal network composed of thermal generating units are used together to meet the total power demand. The scheduling optimization of the hydrothermal power generation system lies in how to reasonably distribute the generation power to each hydroelectric generating unit and thermal generating unit to minimize the total fuel cost of the thermal generating units and minimize the emissions of the thermal generating units, while satisfying all the constraints in the hydro network and the thermal network. Since the power demand changes over time, the power scheduling optimization problem of the hydrothermal power generation system is dynamic. Therefore, when the power demand changes, it is crucial to quickly find a new optimal scheduling scheme.

[0003] There are mainly two existing scheduling optimization methods for hydrothermal power generation systems. The first method is when the power demand changes, a random scheduling scheme is introduced to form a random scheduling scheme population, and the random scheduling scheme population and the optimized scheduling scheme population before the change in power demand are mixed as the initial scheduling scheme population for the current power demand. Then, the initial scheduling scheme population is evolved to obtain the final optimized scheduling scheme population, and a certain scheduling scheme is selected from the optimized scheduling scheme population as the scheduling scheme for the current power demand to perform the scheduling of the hydrothermal power generation system. The second method is when the power demand changes, the scheduling schemes in the optimized scheduling scheme population before the change are updated by crossover and mutation as the initial scheduling scheme population for the current power demand. Then, the initial scheduling scheme population is evolved to obtain the final optimized scheduling scheme population, and a certain scheduling scheme is selected from the optimized scheduling scheme population as the scheduling scheme for the current power demand to perform the scheduling of the hydrothermal power generation system.

[0004] However, the proportion of the random scheduling scheme introduced by the first method after the change in power demand is a fixed value, and the random scheduling scheme has uncertainty and randomness. However, the proportion of the random scheduling scheme introduced by the first method after the change in power demand is a fixed value, and the random scheduling scheme has uncertainty and randomness. Therefore, the initial scheduling scheme population also has relatively large uncertainty and randomness, and there will also be uncertainty in the optimized scheduling scheme population finally obtained by optimizing the initial scheduling scheme population. It may be suitable for the current power demand and has a good effect on balancing the minimization of the total fuel cost of thermal power generation units and the minimization of the emissions of thermal power generation units. It is also possible that it is not suitable for the current power demand and has a poor effect on balancing the minimization of the total fuel cost of thermal power generation units and the minimization of the emissions of thermal power generation units. Therefore, the first method has low stability in terms of the effect of balancing the minimization of the total fuel cost of thermal power generation units and the minimization of the emissions of thermal power generation units for the overall power demand.

[0005] For the second method above, the proportion of crossover and the proportion of mutation for the scheduling schemes in the optimized scheduling scheme population before the change in power demand are both fixed values. The effect of the initial scheduling scheme population obtained through crossover and mutation on minimizing the total fuel cost of thermal power generation units and minimizing the emissions of thermal power generation units highly depends on the values of these two proportions. Therefore, once the values of the crossover and mutation proportions are set inappropriately, it will lead to a significant decline in the effect of balancing the minimization of the total fuel cost of thermal power generation units and the minimization of the emissions of thermal power generation units. The values of the crossover and mutation proportions are set subjectively by humans and have strong subjectivity. They may be suitable or may not be suitable for the current power demand. Therefore, the second method also has low stability in terms of the effect of balancing the minimization of the total fuel cost of thermal power generation units and the minimization of the emissions of thermal power generation units for the overall power demand. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a scheduling optimization method for a hydro-thermal power generation system based on transfer learning and GAN, which has high stability in terms of the effect of balancing the minimization of the total fuel cost of thermal power generation units and the minimization of the emissions of thermal power generation units for the overall power demand.

[0007] The technical solution adopted by the present invention to solve the above technical problems is as follows: A hydrothermal power generation system scheduling optimization method based on transfer learning and GAN. When the current power demand is the first power demand, a random normal distribution method is used to randomly generate multiple scheduling schemes to form an initial scheduling scheme population. Then, each scheduling scheme in the initial scheduling scheme population is respectively subjected to selection, crossover, and mutation operations using the non-dominated sorting genetic algorithm to obtain new scheduling schemes to form an optimized scheduling scheme population for the current power demand. A certain scheduling scheme is selected from the optimized scheduling scheme population for the current power demand as the scheduling scheme for the current power demand to perform hydrothermal power generation system scheduling; when the power demand changes and the changed current power demand is the second power demand, by combining the randomly generated scheduling scheme population with the optimized scheduling scheme population of the previous power demand, an optimized scheduling scheme population for the current power demand is obtained based on the non-dominated sorting genetic algorithm. A certain scheduling scheme is selected from the optimized scheduling scheme population for the current power demand as the scheduling scheme for the current power demand to perform hydrothermal power generation system scheduling; when the power demand changes and the changed current power demand is the third power demand or a power demand after the third power demand, the experience of the historical optimized scheduling scheme population is extracted based on the idea of transfer learning, and a new scheduling scheme population with historical experience is generated by combining the generative adversarial network GAN and the standard normal distribution. An optimized scheduling scheme population for the current power demand is obtained through fusion and the non-dominated sorting genetic algorithm. A certain scheduling scheme is selected from the optimized scheduling scheme population under the current power demand as the scheduling scheme for the current power demand to perform hydrothermal power generation system scheduling.

[0008] Compared with the prior art, the advantages of the present invention are as follows: when the power demand changes and the changed current power demand is the second power demand, the historical optimization scheduling scheme population optimization experience is fully utilized, avoiding the inefficiency of starting from random initialization every time optimization is performed. At the same time, the diversity of the scheduling scheme population is maintained by introducing random schemes, effectively preventing the method from falling into a locally optimal scheduling scheme; when the power demand changes and the changed current power demand is the third power demand or a power demand after the third power demand, the experience of the historical optimization scheduling scheme population is extracted based on the idea of transfer learning, and combined with the generative adversarial network GAN and the standard normal distribution to generate a new scheduling scheme population with historical experience. The optimized scheduling scheme population for the current power demand is obtained through fusion and non-dominated sorting genetic algorithm. By extracting the experience of the historical optimization scheduling scheme population through the idea of transfer learning, the scheduling scheme population information that maintains the historical balance to minimize the total fuel cost of the thermal power generation unit and minimize the emissions of the thermal power generation unit can be retained, and the information of the historical scheduling scheme population can be adaptively adjusted according to the relative change degree of the power demand, no longer strongly relying on artificial parameter settings. The policy information is combined with the generative adversarial network GAN and the standard normal distribution to generate new scheduling schemes. These new scheduling schemes can still maintain the positive introduction of the new scheduling scheme on the basis of maintaining historical information, not being interfered by the locally optimal scheduling scheme. The obtained initial scheduling scheme reduces the severe oscillation during the optimization process, and can maintain the consistency and reliability of the scheduling schemes in the optimized scheduling scheme population under the current power demand obtained during the continuous power demand change process, improving the stability of the effect of balancing the minimization of the total fuel cost of the thermal power generation unit and the minimization of the emissions of the thermal power generation unit. Therefore, the present invention can adaptively utilize the scheduling scheme population of historical power demands according to the relative change degree of the power demand, so as to generate a better initial scheduling scheme population for the current power demand, and has a relatively high stability in the effect of balancing the minimization of the total fuel cost of the thermal power generation unit and the minimization of the emissions of the thermal power generation unit for the overall power demand.

[0009] Further, when the power demand changes and the current power demand after the change is the second power demand, the specific process of obtaining the optimized scheduling plan population for the current power demand based on the non-dominated sorting genetic algorithm by combining the randomly generated scheduling plan population and the optimized scheduling plan population for the previous power demand is as follows: First, use the random normal distribution method to randomly generate multiple scheduling plans to form a random scheduling plan population. Mix the currently obtained random scheduling plan population and the optimized scheduling plan population for the first power demand to form a temporary scheduling plan population. Then, based on the non-dominated sorting operation, select the required number of scheduling plans from the temporary scheduling plan population to form the initial scheduling plan population for the current power demand. Next, perform selection, crossover, and mutation operations on each scheduling plan in the initial scheduling plan population for the current power demand using the non-dominated sorting genetic algorithm to obtain new scheduling plans to form the optimized scheduling plan population for the current power demand.

[0010] Further, when the power demand changes and the current power demand after the change is the third power demand or a power demand after the third power demand, the specific process of extracting the experience of the historical optimized scheduling plan population based on the idea of transfer learning, combining the generative adversarial network GAN and the standard normal distribution to generate a new scheduling plan population with historical experience, and obtaining the optimized scheduling plan population for the current power demand through fusion and the non-dominated sorting genetic algorithm is as follows:

[0011] Step 1: Perform non-dominated sorting on the scheduling plans in the optimized scheduling plan population for the previous power demand, and divide all the scheduling plans in the optimized scheduling plan population for the previous power demand into different levels. After the division, each scheduling plan in the first level is better than each scheduling plan in the second level in terms of minimizing the total combustion cost of the thermal power generation units and minimizing the emissions of the thermal power generation units. Each scheduling plan in the second level is better than each scheduling plan in the third level in terms of minimizing the total combustion cost of the thermal power generation units and minimizing the emissions of the thermal power generation units, and so on. That is, each scheduling plan in a lower level is better than each scheduling plan in a higher level in terms of minimizing the total combustion cost of the thermal power generation units and minimizing the emissions of the thermal power generation units.

[0012] Step 2: Calculate the total fuel cost of the thermal power generation units and the emissions of the thermal power generation units for each dispatching scheme at the first level. Use the total fuel cost of the thermal power generation units as the abscissa of the two-dimensional coordinate system and the emissions of the thermal power generation units as the ordinate of the two-dimensional coordinate system to form a two-dimensional coordinate system. The total fuel cost of the thermal power generation units for each dispatching scheme at the first level is used as its abscissa, and the emissions of the thermal power generation units of the thermal power generation point units are used as its ordinate to form its two-dimensional coordinate point. Mark the two-dimensional coordinate points of each dispatching scheme at the first level in the two-dimensional coordinate system. Use the minimum value of the abscissas of the two-dimensional coordinate points of all dispatching schemes at the first level as the abscissa and the maximum value of the ordinates as the ordinate to form the first boundary coordinate, which is marked as boundary point one in the two-dimensional coordinate system. Use the maximum value of the abscissas among the two-dimensional coordinate points of all dispatching schemes at the first level as the abscissa and the minimum value of the ordinates as the ordinate to form the second boundary coordinate, which is marked as boundary point two in the two-dimensional coordinate system. Connect boundary point one and boundary point two with a straight line, and use the line segment between boundary point one and boundary point two as the extreme value line;

[0013] Step 3: Segment the horizontal axis between the abscissa of boundary point one and the abscissa of boundary point two in the two-dimensional coordinate system to obtain multiple horizontal axis segments. The specific process is as follows: Set an integer greater than or equal to 1 as the number of segments. If the length value of each segment obtained by evenly segmenting this horizontal axis according to the number of segments is an integer or a decimal with no more than two digits after the decimal point, then evenly segment the horizontal axis between the abscissa of boundary point one and the abscissa of boundary point two according to the number of segments to obtain multiple horizontal axis segments. If the length value of each segment obtained is a decimal with more than two digits after the decimal point, then use the truncation method to retain two digits after the decimal point for this length value to obtain the truncated length value. Starting from the abscissa of boundary point one, segment this horizontal axis with the truncated length value until the length of the remaining horizontal axis segment is less than the truncated length value, and use it as the last horizontal axis segment to obtain multiple horizontal axis segments;

[0014] Step 4: Draw straight lines perpendicular to the horizontal axis of the two-dimensional coordinate system at the two endpoints of each horizontal axis segment. The straight lines perpendicular to the horizontal axis of the two-dimensional coordinate system drawn at the two endpoints of each horizontal axis segment are called its two perpendicular lines. The two perpendicular lines of each horizontal axis segment will intersect the extreme value line. A quadrilateral region is formed between the two perpendicular lines of each horizontal axis segment, the extreme value line, and the horizontal axis of the two-dimensional coordinate system, obtaining multiple quadrilateral regions. At this time, the two-dimensional coordinate points of all dispatching schemes at the first level are divided into these quadrilateral regions;

[0015] Step 5: Take the direction from the abscissa of boundary point 1 to the abscissa of boundary point 2 as the left-to-right direction. If the two-dimensional coordinate point of a scheduling scheme at the first level falls on the vertical line drawn from the left endpoint of the leftmost horizontal axis segment, it is considered to be in the leftmost quadrilateral region. If the two-dimensional coordinate point of a scheduling scheme at the first level falls on the vertical line drawn from the right endpoint of the rightmost horizontal axis segment, it is considered to be in the rightmost quadrilateral region. If the two-dimensional coordinate point of a scheduling scheme falls on other vertical lines except the vertical line drawn from the left endpoint of the leftmost horizontal axis segment and the vertical line drawn from the right endpoint of the rightmost horizontal axis segment, it is considered to be in the quadrilateral region with this vertical line as the left side;

[0016] Step 6: Calculate the perpendicular distance from the two-dimensional coordinate point of each scheduling scheme at the first level to the extreme value line. Select the two-dimensional coordinate point with the largest perpendicular distance from all two-dimensional coordinate points in each quadrilateral region to the extreme value line as an alternative two-dimensional coordinate point. If there are multiple, randomly select one of them. The scheduling schemes corresponding to all the alternative two-dimensional coordinate points obtained at this time form the first scheduling scheme source population;

[0017] Step 7: Calculate the mutual information values between each scheduling scheme in the optimized scheduling scheme population of the previous power demand of the current power demand and each scheduling scheme in the optimized scheduling scheme population of the penultimate power demand of the current power demand. Use all the calculated mutual information values to form a mutual information matrix. Each row of the mutual information matrix represents the mutual information values between a scheduling scheme in the optimized scheduling scheme of the previous power demand and all the scheduling schemes in the optimized scheduling scheme population of the penultimate power demand of the current power demand. Each element in each column of each row of the mutual information matrix represents the mutual information value between a scheduling scheme in the optimized scheduling scheme of the previous power demand and a scheduling scheme in the optimized scheduling scheme population of the penultimate power demand of the current power demand. Each column of the mutual information matrix represents the mutual information values between each scheduling scheme in the optimized scheduling scheme of the previous power demand and a certain scheduling scheme in the optimized scheduling scheme population of the penultimate power demand of the current power demand. Each element in each row of each column of the mutual information matrix represents the mutual information value between a scheduling scheme in the optimized scheduling scheme of the penultimate power demand and a scheduling scheme in the optimized scheduling scheme population of the previous power demand of the current power demand;

[0018] Step 8: Use the mutual information value between a certain scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand and a certain scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand as the preference value between the two. For each scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand, sort the mutual information values between it and all scheduling plans in the optimal scheduling plan population of the penultimate power demand of the current power demand from high to low to form its preference list, so as to obtain the preference list of each scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand. For each scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand, sort the mutual information values between it and all scheduling plans in the optimal scheduling plan population of the previous power demand of the current power demand from high to low to form its preference list, so as to obtain the preference list of each scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand;

[0019] Step 9: According to the preference list of each scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand and the preference list of each scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand, use the stable marriage matching strategy to match between each scheduling plan in the optimal scheduling plan population of the penultimate power demand and each scheduling plan in the optimal scheduling plan population of the previous power demand, so that the scheduling plans in the optimal scheduling plan population of the previous power demand of the current power demand are in one-to-one correspondence with the scheduling plans in the optimal scheduling plan population of the penultimate power demand of the current power demand, and obtain multiple stable matching scheduling plan pairs. Each stable matching scheduling plan pair is composed of a scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand and a scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand;

[0020] Step 10: Perform non-dominated sorting on the population of optimal scheduling solutions for the previous power demand of the current power demand, and divide all the scheduling solutions in the population of optimal scheduling solutions for the previous power demand into different levels. After division, each scheduling solution in the first level is better than each scheduling solution in the second level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units. Each scheduling solution in the second level is better than each scheduling solution in the third level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units, and so on. That is, each scheduling solution in a lower level is better than each scheduling solution in a higher level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units. Sort the scheduling solutions in each level from the first level to the last level to obtain the sorting of all the scheduling solutions in the population of optimal scheduling solutions for the previous power demand of the current power demand. According to the sorting of all the scheduling solutions in the population of optimal scheduling solutions for the previous power demand of the current power demand, sort the obtained multiple stable matching scheduling solution pairs to obtain the sorted stable matching scheduling solution pairs, and then select the first 50% of the stable matching scheduling solution pairs from the sorted stable matching scheduling solution pairs. If 50% of the total number of stable matching scheduling solution pairs is not an integer, round up. Use the scheduling solutions in all the selected stable matching scheduling solution pairs to form the scheduling solution source population two.

[0021] Step 11: Merge the scheduling solution source population one and the scheduling solution source population two to form the scheduling solution source population three.

[0022] Step 12: Calculate the mean of the scheduling solution source population one, the mean of the scheduling solution source population two, the total combustion cost of the thermal power generation units and the emissions of the thermal power generation units of each scheduling solution in the scheduling solution source population one for the current power demand, and the total combustion cost of the thermal power generation units and the emissions of the thermal power generation units of each scheduling solution in the scheduling solution source population one for the previous power demand of the current power demand respectively. Obtain the relative change degree between the current power demand and the previous power demand according to the relative change amount between the total combustion cost of the thermal power generation units and the emissions of the thermal power generation units of each scheduling solution in the scheduling solution source population one for the current power demand and the total combustion cost of the thermal power generation units and the emissions of the thermal power generation units of the previous power demand. The value range of the relative change degree is [0, 1]. Use the relative change degree to perform weighted summation on the mean of the scheduling solution source population one and the mean of the scheduling solution source population two respectively, and use the weighted summation result obtained at this time as the adaptive comprehensive mean of the scheduling solution source population three. Then calculate the covariance of the scheduling solution source three.

[0023] Step 13: Set the input of the generator of the generative adversarial network (GAN) to a multivariate Gaussian distribution that conforms to the adaptive comprehensive mean and covariance of the source population three of the scheduling scheme. Use each scheduling scheme in the optimized scheduling scheme population of the previous electricity demand as the input of the discriminator of the GAN, perform adversarial training on the generator and discriminator of the GAN to obtain the trained GAN, and use the trained GAN to generate the source population four of the scheduling scheme;

[0024] Step 14: Perform Cholesky decomposition on the covariance matrix of the source population three of the scheduling scheme to obtain a lower triangular matrix. Randomly generate multiple scheduling schemes using the standard normal distribution method, multiply each generated scheduling scheme by the lower triangular matrix respectively to obtain a scheduling scheme with a specified covariance matrix corresponding to each scheduling scheme, add the adaptive comprehensive mean of the source population three of the scheduling scheme to each currently obtained scheduling scheme with a specified covariance matrix respectively to obtain a new scheduling scheme, and use these new scheduling schemes to form the source population five of the scheduling scheme;

[0025] Step 15: Merge the source population four of the scheduling scheme, the source population five of the scheduling scheme, and the optimized scheduling scheme population of the previous electricity demand to form the source population six of the scheduling scheme;

[0026] Step 16: Use non-dominated sorting operation to select the required number of scheduling schemes from the source population six of the scheduling scheme to form the initial scheduling scheme population under the current electricity demand, and then perform selection, crossover, and mutation operations on the initial scheduling scheme population under the current electricity demand using the non-dominated sorting genetic algorithm to obtain new scheduling schemes to form the optimized scheduling scheme population of the current electricity demand.

[0027] Furthermore, the mean of the source population one of the scheduling scheme and the mean of the source population two of the scheduling scheme are calculated using formulas (1) and (2) respectively:

[0028]

[0029] where μ k is the mean of the source population one of the scheduling scheme, S is the total number of scheduling schemes in the source population one of the scheduling scheme, kind i is the i-th scheduling scheme in the source population one of the scheduling scheme, i = 1, 2,..., S, μ r is the mean of the source population two of the scheduling scheme, Z is the number of scheduling schemes in the source population two of the scheduling scheme, rind j is the j-th scheduling scheme in the source population two of the scheduling scheme, j = 1, 2,..., Z.

[0030] Furthermore, the relative change degree between the current electricity demand and the previous electricity demand of the current electricity demand is calculated using formula (3):

[0031]

[0032] Among them, M is the number of objective functions in the mathematical model obtained after digital-analog modeling of the hydrothermal power generation scheduling system, v = 1, 2, …, M, u = 1, 2, …, S, f v (x u ) is the function value of the v-th objective function of the u-th scheduling plan x u in the first scheduling plan source population obtained through the mathematical model under the current power demand, is the function value of the v-th objective function of the u-th scheduling plan x u in the previous power demand of the current power demand obtained through the mathematical model, Max v is the maximum function value among the function values of the v-th objective function of all scheduling plans in the first scheduling plan source population obtained through the mathematical model, Min v is the minimum function value among the function values of the v-th objective function of all scheduling plans in the first scheduling plan source population obtained through the mathematical model, δ is the relative change degree in the range of [0, 1] between the current power demand and the previous power demand of the current power demand, and || is the absolute value symbol.

[0033] Furthermore, the adaptive comprehensive mean of the third scheduling plan source population is calculated using formula (4):

[0034] MV = δμ k + (1 - δ)μ r (4)

[0035] where MV is the adaptive comprehensive mean of the third scheduling plan source population.

[0036] Further, the generator of the generative adversarial network (GAN) adopts a multi-layer feedforward neural network. The generator includes two hidden layers and an output layer. The first hidden layer of the generator is configured with 512 neurons, and the second hidden layer of the generator is configured with 256 neurons. The first hidden layer of the generator adopts the LeakyReLU activation function, and the second hidden layer of the generator uses the ELU activation function. The output layer of the generator outputs the scheduling scheme generated by the generator through the Tanh activation function. The input of the generator is the scheduling scheme sampled from the multivariate Gaussian distribution that conforms to the adaptive comprehensive mean and covariance of the scheduling scheme source population three. The discriminator of the generative adversarial network (GAN) adopts a lightweight fully connected network structure. The discriminator of the generative adversarial network (GAN) includes a hidden layer and an output layer. The hidden layer of the discriminator is configured with 256 neurons and uses Swish as the activation function. The input of the discriminator is each scheduling scheme in the optimized scheduling scheme population of the previous electricity demand and each scheduling scheme generated by the generator. Each scheduling scheme in the optimized scheduling scheme population of the previous electricity demand serves as the real dataset of the discriminator, and each scheduling scheme generated by the generator serves as the generated dataset of the discriminator. The output layer of the discriminator adopts the Sigmoid activation function to obtain the probability that each scheduling scheme in the real dataset and the generated dataset is a real scheduling scheme. When performing adversarial training on the generator and discriminator of the generative adversarial network (GAN), the batch size is set to 32, the number of training epochs is set to 100, and the gradient descent method is used to update the generator and discriminator of the GAN during the adversarial training process. Description of the Drawings

[0037] Figure 1 Flowchart of the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention when the current electricity demand is the first electricity demand;

[0038] Figure 2 Flowchart of the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention when the current electricity demand is the second electricity demand;

[0039] Figure 3 Flowchart of the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention when the current electricity demand is the third or more electricity demands;

[0040] Figure 4 Data graph of the total combustion cost of thermal power generating units and the emissions of thermal power generating units under different power generation demands obtained from experiments on the hydrothermal power generation scheduling system for the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention;

[0041] Figure 5 is a comparison chart of algorithm performance under specific power generation requirements obtained from experiments on the hydrothermal power generation scheduling system using the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention;

[0042] Figure 6 is a comparison chart of algorithm performance at different power generation demand levels obtained from experiments on the hydrothermal power generation scheduling system using the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention;

[0043] Figure 7 It is a data chart of the percentage improvement in average emissions under different demands obtained from experiments on the hydrothermal power generation scheduling system using the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention. Detailed implementation manners

[0044] The present invention will be further described in detail below in conjunction with the embodiments with reference to the drawings.

[0045] Embodiment 1: A hydrothermal power generation system scheduling optimization method based on transfer learning and GAN. When the current power demand is the first power demand, a plurality of scheduling schemes are randomly generated using the random normal distribution method to form an initial scheduling scheme population. Then, each scheduling scheme in the initial scheduling scheme population is respectively subjected to selection, crossover, and mutation operations using the non-dominated sorting genetic algorithm to obtain new scheduling schemes to form an optimized scheduling scheme population for the current power demand. A certain scheduling scheme is selected from the optimized scheduling scheme population for the current power demand as the scheduling scheme for the current power demand to perform hydrothermal power generation system scheduling; when the power demand changes and the changed current power demand is the second power demand, by combining the randomly generated scheduling scheme population with the optimized scheduling scheme population for the previous power demand, an optimized scheduling scheme population for the current power demand is obtained based on the non-dominated sorting genetic algorithm. A certain scheduling scheme is selected from the optimized scheduling scheme population for the current power demand as the scheduling scheme for the current power demand to perform hydrothermal power generation system scheduling; when the power demand changes and the changed current power demand is the third power demand or a power demand after the third power demand, the experience of the historical optimized scheduling scheme population is extracted based on the idea of transfer learning, and a new scheduling scheme population with historical experience is generated by combining the generative adversarial network GAN and the standard normal distribution. An optimized scheduling scheme population for the current power demand is obtained through fusion and the non-dominated sorting genetic algorithm. A certain scheduling scheme is selected from the optimized scheduling scheme population under the current power demand as the scheduling scheme for the current power demand to perform hydrothermal power generation system scheduling.

[0046] Embodiment 2: This embodiment is basically the same as Embodiment 1, the difference being that: in this embodiment, as Figure 2As shown, when the power demand changes and the current power demand after the change is the second power demand, the specific process of obtaining the optimized scheduling plan population for the current power demand by combining the randomly generated scheduling plan population and the optimized scheduling plan population of the previous power demand based on the non-dominated sorting genetic algorithm is as follows: First, use the random normal distribution method to randomly generate multiple scheduling plans to form a random scheduling plan population, mix the currently obtained random scheduling plan population and the optimized scheduling plan population of the first power demand to form a temporary scheduling plan population, and then select the required number of scheduling plans from the temporary scheduling plan population based on the non-dominated sorting operation to form the initial scheduling plan population for the current power demand. Then, perform selection, crossover, and mutation operations on each scheduling plan in the initial scheduling plan population for the current power demand using the non-dominated sorting genetic algorithm to obtain new scheduling plans to form the optimized scheduling plan population for the current power demand.

[0047] In this embodiment, when the power demand changes and the current power demand after the change is the second power demand, during the optimization process, the optimization experience of the historical optimized scheduling plan population is fully utilized, avoiding the inefficiency of starting from random initialization every time for optimization. At the same time, the diversity of the scheduling plan population is maintained by the introduction of random plans, effectively preventing the method from falling into a locally optimal scheduling plan. Thus, when determining the scheduling plan for the second power demand, a high-quality initial scheduling plan population can be formed, improving the optimization efficiency and the reliability of the optimized scheduling plan population for the current power demand.

[0048] Embodiment 3: This embodiment is basically the same as Embodiment 2, except that: in this embodiment, as Figure 3 shown, when the power demand changes and the current power demand after the change is the third power demand or the power demand after the third power demand, extract the experience of the historical optimized scheduling plan population based on the idea of transfer learning, and combine the generative adversarial network GAN and the standard normal distribution to generate a new scheduling plan population with historical experience. The specific process of obtaining the optimized scheduling plan population for the current power demand through fusion and the non-dominated sorting genetic algorithm is as follows:

[0049] Step 1: Perform non-dominated sorting on the scheduling plans in the population of the optimization scheduling plan for the previous electricity demand, and divide all the scheduling plans in the population of the optimization scheduling plan for the previous electricity demand into different levels. After division, each scheduling plan in the first level is better than each scheduling plan in the second level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units. Each scheduling plan in the second level is better than each scheduling plan in the third level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units, and so on. That is, each scheduling plan in a lower level is better than each scheduling plan in a higher level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units;

[0050] Step 2: Calculate the total fuel cost of the thermal power generation units and the emissions of the thermal power generation units for each scheduling plan in the first level. Use the total fuel cost of the thermal power generation units as the abscissa of the two-dimensional coordinate system and the emissions of the thermal power generation units as the ordinate of the two-dimensional coordinate system to form a two-dimensional coordinate system. The total fuel cost of the thermal power generation units of each scheduling plan in the first level is used as its abscissa, and the emissions of the thermal power generation units are used as its ordinate to form its two-dimensional coordinate point. Mark the two-dimensional coordinate points of each scheduling plan in the first level in the two-dimensional coordinate system. Use the minimum value of the abscissas of the two-dimensional coordinate points of all scheduling plans in the first level as the abscissa and the maximum value of the ordinates as the ordinate to form the first boundary coordinate, which is marked as boundary point one in the two-dimensional coordinate system. Use the maximum value of the abscissas of the two-dimensional coordinate points of all scheduling plans in the first level as the abscissa and the minimum value of the ordinates as the ordinate to form the second boundary coordinate, which is marked as boundary point two in the two-dimensional coordinate system. Connect boundary point one and boundary point two with a straight line, and use the line segment between boundary point one and boundary point two as the extreme value line;

[0051] Step 3: Segment the horizontal axis between the abscissa of boundary point one and the abscissa of boundary point two in the two-dimensional coordinate system to obtain multiple horizontal axis segments. The specific process is as follows: Set an integer greater than or equal to 1 as the number of segments. If the length value of each segment obtained by evenly segmenting this horizontal axis according to the number of segments is an integer or a decimal with no more than two digits after the decimal point, then evenly segment the horizontal axis between the abscissa of boundary point one and the abscissa of boundary point two according to the number of segments to obtain multiple horizontal axis segments. If the length value of each segment obtained is a decimal with more than two digits after the decimal point, then use the truncation method to retain two digits after the decimal point for this length value to obtain the truncated segment length value. Starting from the abscissa of boundary point one, segment this horizontal axis with the truncated segment length value until the length of the remaining horizontal axis segment is less than the truncated segment length value, and use it as the last horizontal axis segment to obtain multiple horizontal axis segments;

[0052] Step 4: Straight lines perpendicular to the horizontal axis of the two-dimensional coordinate system are drawn at the two endpoints of each horizontal axis segment. The two straight lines perpendicular to the horizontal axis of the two-dimensional coordinate system drawn at the two endpoints of each horizontal axis segment are called its two perpendicular lines. The two perpendicular lines of each horizontal axis segment will intersect the extreme value line. A quadrilateral region is formed among the two perpendicular lines of each horizontal axis segment, the extreme value line, and the horizontal axis of the two-dimensional coordinate system, obtaining multiple quadrilateral regions. At this time, the two-dimensional coordinate points of all scheduling schemes at the first level are divided into these quadrilateral regions;

[0053] Step 5: The direction from the abscissa of boundary point 1 to the abscissa of boundary point 2 is taken as the left-to-right direction; if the two-dimensional coordinate point of a certain scheduling scheme at the first level falls on the perpendicular line drawn at the left endpoint of the leftmost horizontal axis segment, it is considered to be in the leftmost quadrilateral region. If the two-dimensional coordinate point of a certain scheduling scheme at the first level falls on the perpendicular line drawn at the right endpoint of the rightmost horizontal axis segment, it is considered to be in the rightmost quadrilateral region. If the two-dimensional coordinate point of a certain scheduling scheme falls on other perpendicular lines except the perpendicular line drawn at the left endpoint of the leftmost horizontal axis segment and the perpendicular line drawn at the right endpoint of the rightmost horizontal axis segment, it is considered to be in the quadrilateral region with this perpendicular line as the left side;

[0054] Step 6: Calculate the perpendicular line distance from the two-dimensional coordinate point of each scheduling scheme at the first level to the extreme value line. Select the two-dimensional coordinate point with the largest perpendicular line distance from all two-dimensional coordinate points in each quadrilateral region to the extreme value line as an alternative two-dimensional coordinate point. If there are multiple, randomly select one of them. The scheduling schemes corresponding to all the alternative two-dimensional coordinate points obtained at this time form the first source population of scheduling schemes;

[0055] Step 7: Calculate the mutual information values between each scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand and each scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand. Use all the calculated mutual information values to form a mutual information matrix. Each row in the mutual information matrix represents the mutual information values between a scheduling plan in the optimal scheduling plan of the previous power demand and all the scheduling plans in the optimal scheduling plan population of the penultimate power demand of the current power demand. Each element in each column of each row in the mutual information matrix represents the mutual information value between a scheduling plan in the optimal scheduling plan of the previous power demand and a scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand. Each column in the mutual information matrix represents the mutual information values between each scheduling plan in the optimal scheduling plan of the previous power demand and a certain scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand. Each element in each row of each column in the mutual information matrix represents the mutual information value between a scheduling plan in the optimal scheduling plan of the penultimate power demand and a scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand.

[0056] Step 8: Take the mutual information value between a certain scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand and a certain scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand as the preference value between the two. For each scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand, sort the mutual information values between it and all the scheduling plans in the optimal scheduling plan population of the penultimate power demand of the current power demand from high to low to form its preference list, so as to obtain the preference list of each scheduling plan in the optimal scheduling plan population of the previous power demand of the current power demand. For each scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand, sort the mutual information values between it and all the scheduling plans in the optimal scheduling plan population of the previous power demand of the current power demand from high to low to form its preference list, so as to obtain the preference list of each scheduling plan in the optimal scheduling plan population of the penultimate power demand of the current power demand.

[0057] Step 9: According to the preference list of each scheduling plan in the optimized scheduling plan population of the previous electricity demand for the current electricity demand and the preference list of each scheduling plan in the optimized scheduling plan population of the penultimate electricity demand for the current electricity demand, use the stable marriage matching strategy to match each scheduling plan in the optimized scheduling plan population of the penultimate electricity demand with each scheduling plan in the optimized scheduling plan population of the previous electricity demand, so that the scheduling plans in the optimized scheduling plan population of the previous electricity demand for the current electricity demand are in one-to-one correspondence with the scheduling plans in the optimized scheduling plan population of the penultimate electricity demand for the current electricity demand, and obtain multiple stable matching scheduling plan pairs. Each stable matching scheduling plan pair consists of a scheduling plan in the optimized scheduling plan population of the previous electricity demand for the current electricity demand and a scheduling plan in the optimized scheduling plan population of the penultimate electricity demand for the current electricity demand;

[0058] Step 10: Perform non-dominated sorting on the optimized scheduling plan population of the previous electricity demand for the current electricity demand, and divide all the scheduling plans in the optimized scheduling plan population of the previous electricity demand into different levels. After division, each scheduling plan in the first level is better than each scheduling plan in the second level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units. Each scheduling plan in the second level is better than each scheduling plan in the third level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units, and so on. That is, each scheduling plan in a lower level is better than each scheduling plan in a higher level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units; Sort the scheduling plans in each level from the first level to the last level to obtain the sorting of all the scheduling plans in the optimized scheduling plan population of the previous electricity demand for the current electricity demand; According to the sorting of all the scheduling plans in the optimized scheduling plan population of the previous electricity demand for the current electricity demand, sort the obtained multiple stable matching scheduling plan pairs to obtain the sorted stable matching scheduling plan pairs, and then select the first 50% of the stable matching scheduling plan pairs from the sorted stable matching scheduling plan pairs. If 50% of the total number of stable matching scheduling plan pairs is not an integer, round up. Use the scheduling plans in all the selected stable matching scheduling plan pairs to form the scheduling plan source population two;

[0059] Step 11: Merge the scheduling plan source population one and the scheduling plan source population two to form the scheduling plan source population three;

[0060] Step 12: Calculate the mean of the source population one of the scheduling plans, the mean of the source population two of the scheduling plans, the total combustion cost and emissions of the thermal power generators for each scheduling plan in the source population one of the scheduling plans under the current power demand, and the total combustion cost and emissions of the thermal power generators for each scheduling plan in the source population one of the scheduling plans under the previous power demand of the current power demand. Obtain the relative change degree between the current power demand and the previous power demand according to the relative change amount between the total combustion cost and emissions of the thermal power generators for each scheduling plan in the source population one of the scheduling plans under the current power demand and the total combustion cost and emissions of the thermal power generators for each scheduling plan in the source population one of the scheduling plans under the previous power demand. The value range of the relative change degree is [0, 1]. Use the relative change degree to perform weighted summation on the mean of the source population one of the scheduling plans and the mean of the source population two of the scheduling plans respectively, and take the weighted summation result obtained at this time as the adaptive comprehensive mean of the source population three of the scheduling plans, and then calculate the covariance of the source population three of the scheduling plans;

[0061] Step 13: Set the input of the generator of the generative adversarial network GAN to a multivariate Gaussian distribution that conforms to the adaptive comprehensive mean and covariance of the source population three of the scheduling plans, and use each scheduling plan in the optimized scheduling plan population of the previous power demand as the input of the discriminator of the GAN. Perform adversarial training on the generator and discriminator of the GAN to obtain the trained GAN, and use the trained GAN to generate the source population four of the scheduling plans;

[0062] Step 14: Perform Cholesky decomposition on the covariance matrix of the source population three of the scheduling plans to obtain a lower triangular matrix. Randomly generate multiple scheduling plans using the standard normal distribution method. Multiply each scheduling plan generated at this time by the lower triangular matrix respectively to obtain a scheduling plan corresponding to each scheduling plan with a specified covariance matrix. Add the adaptive comprehensive mean of the source population three of the scheduling plans to each scheduling plan with a specified covariance matrix obtained currently to obtain a new scheduling plan. Use these new scheduling plans to form the source population five of the scheduling plans;

[0063] Step 15: Merge the source population four of the scheduling plans, the source population five of the scheduling plans, and the optimized scheduling plan population of the previous power demand to form the source population six of the scheduling plans;

[0064] Step 16: Use non-dominated sorting operation to select the required number of scheduling plans from the source population six of the scheduling plans to form the initial scheduling plan population under the current power demand. Then, perform selection, crossover, and mutation operations on the initial scheduling plan population under the current power demand using the non-dominated sorting genetic algorithm to obtain new scheduling plans to form the optimized scheduling plan population of the current power demand.

[0065] In this embodiment, when the power demand changes and the current power demand after the change is the third power demand or the power demand after the third power demand, an initial scheduling plan population for the current power demand is obtained by combining transfer learning and generative adversarial networks. This not only considers the current power demand but also makes full use of historical optimization experience, supplemented by mutual information and stable marriage matching strategies to establish the association between historical scheduling plans. Moreover, an adaptive comprehensive mean calculation method that weights and fuses the mean of the historical optimized scheduling plan population according to the degree of power demand change enhances the adaptability of the scheduling plan to power demand changes, constructs multiple sources of scheduling plan populations, and adopts the fusion of multiple sources and non-dominated sorting methods, which not only ensures the effect of the optimized scheduling plan in balancing the total fuel cost of thermal power generation units to a minimum and minimizing the emissions of thermal power generation units but also enhances the diversity of the scheduling plan.

[0066] Embodiment 4: This embodiment is basically the same as Embodiment 2, except that: in this embodiment, the means of the scheduling plan source population 1 and the scheduling plan source population 2 are calculated by formulas (1) and (2) respectively:

[0067]

[0068] where μ k is the mean of the scheduling plan source population 1, S is the total number of scheduling plans in the scheduling plan source population 1, kind i is the i-th scheduling plan in the scheduling plan source population 1, i = 1, 2,..., S, μ r is the mean of the scheduling plan source population 2, Z is the number of scheduling plans in the scheduling plan source population 2, rind j is the j-th scheduling plan in the scheduling plan source population 2, j = 1, 2,..., Z.

[0069] In this embodiment, the relative change degree between the current power demand and the previous power demand of the current power demand is calculated by formula (3):

[0070]

[0071] where M is the number of objective functions in the mathematical model obtained after digital simulation modeling of the hydrothermal power generation scheduling system, v = 1, 2,..., M, u = 1, 2,..., S, f v (x u ) is the function value of the v-th objective function of the u-th scheduling plan x u in the current power demand obtained by the u-th scheduling plan in the scheduling plan source population 1 through the mathematical model, is the u-th scheduling plan x uThe function value of the v-th objective function of the previous power demand obtained through the mathematical model, Max v is the maximum function value among the function values of the v-th objective function of all scheduling plans in the source population one of the scheduling plans obtained through the mathematical model for the current power demand, Min v is the minimum function value among the function values of the v-th objective function of all scheduling plans in the source population one of the scheduling plans obtained through the mathematical model for the current power demand. δ is the relative change degree in the range of [0, 1] between the current power demand and the previous power demand of the current power demand, and || is the absolute value symbol.

[0072] In this embodiment, the adaptive comprehensive mean of the source population three of the scheduling plans is calculated using formula (4):

[0073] MV = δμ k +(1 - δ)μ r (4)

[0074] where MV is the adaptive comprehensive mean of the source population three of the scheduling plans.

[0075] In this embodiment, the generator of the generative adversarial network GAN adopts a multi-layer feedforward neural network. The generator includes two hidden layers and an output layer. The first hidden layer of the generator is configured with 512 neurons, the second hidden layer of the generator is configured with 256 neurons. The first hidden layer of the generator adopts the LeakyReLU activation function, the second hidden layer of the generator uses the ELU activation function. The output layer of the generator outputs the scheduling plan generated by the generator through the Tanh activation function. The input of the generator is the scheduling plan sampled from the multivariate Gaussian distribution that conforms to the adaptive comprehensive mean and covariance of the source population three of the scheduling plans. The discriminator of the generative adversarial network GAN adopts a lightweight fully connected network structure. The discriminator of the generative adversarial network GAN includes a hidden layer and an output layer. The hidden layer of the discriminator is configured with 256 neurons and adopts Swish as the activation function. The input of the discriminator is each scheduling plan in the optimized scheduling plan population of the previous power demand and each scheduling plan generated by the generator. Each scheduling plan in the optimized scheduling plan population of the previous power demand serves as the real dataset of the discriminator, and each scheduling plan generated by the generator serves as the generated dataset of the discriminator. The output layer of the discriminator adopts the Sigmoid activation function to obtain the probability that each scheduling plan in the real dataset and the generated dataset is a real scheduling plan; when performing adversarial training on the generator and discriminator of the generative adversarial network GAN, the batch size is set to 32, the training period is set to 100, and the gradient descent method is used to update the generator and discriminator of the GAN during the adversarial training process.

[0076] In this embodiment, by quantifying the relative change degree between power demands, this quantified index is used as a key weight factor to dynamically adjust the mean of the historical optimal scheduling plan population, thereby constructing an adaptive comprehensive mean, enabling the scheduling plan to intelligently perceive the change degree of power demands and then adjust the historical experience reference preference of the mean. The multivariate Gaussian distribution that conforms to the adaptive comprehensive mean and covariance is used as the input of the GAN, making full use of the historical optimization experience. The newly generated scheduling plan has both randomness and maintains similarity with the historical high-quality scheduling plans, avoiding a large number of unreasonable scheduling plans that may be generated by random methods. At the same time, the optimal scheduling plan population of the previous power demand is used as the real data set of the discriminator, ensuring that the newly generated plan is close to the distribution characteristics of the historical scheduling plans. Finally, the integration of the scheduling plan population generated by the GAN, the scheduling plan population generated by the standard normal distribution, and the optimal scheduling plan population of the previous power demand constitutes a comprehensive scheduling plan population with both historical experience inheritance and innovative introduction, providing a high-quality initial scheduling plan population space for the subsequent non-dominated genetic algorithm.

[0077] To verify the performance of the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention, in the hardware environment equipped with a 12th Gen Intel(R) Core(TM) i7-12700H processor and an NVIDIA GeForce RTX3060 Laptop GPU, the MATLAB R2022a software environment is used to implement the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention, and the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention is simulated.

[0078] During the simulation process, the optimization objectives of the mathematical modeling of the hydrothermal power generation scheduling system are set to minimize the total fuel cost of the thermal power generation units and minimize the emissions of the thermal power generation units. Therefore, the number of objective functions M is 2. A hydrothermal network composed of two hydroelectric generating units and a thermal network composed of four thermal power generation units are set. The number of power demand changes is 4. The number of scheduling plans in each optimal scheduling plan population of each power demand is 100, the dimension of each scheduling plan is 24, and the crossover and mutation probabilities of the non-dominated genetic algorithm are 0.9 and 0.01 respectively.

[0079] After the simulation experiment is completed, according to the simulation experiment results of the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN (abbreviated as GTLP) of the present invention, Figure 4 the result diagrams of the total fuel cost of the thermal power generation units and the emissions of the thermal power generation units of each optimal scheduling plan population of the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention under four power demands as shown are drawn. Figure 4In this figure, the horizontal axis represents the total fuel cost of the thermal power generating units in the scheduling plan, and the vertical axis represents the emissions of the thermal power generating units.

[0080] To compare with the hydro-thermal power system scheduling optimization method based on transfer learning and GAN of the present invention, under the same environment and parameter settings, the hydro-thermal power system scheduling optimization method using the dynamic non-dominated genetic algorithm disclosed in the literature "Dynamic Multi-Objective Optimization and Decision-Making Using Modified NSGA-II: A Case Study on Hydro-Thermal Power Scheduling" (abbreviated as DNSGA-II-A) is implemented, and simulation experiments are carried out on it. After the simulation experiments are completed, according to the simulation experiment results of DNSGA-II-A and the hydro-thermal power system scheduling optimization method based on transfer learning and GAN of the present invention, the performance comparison diagrams of DNSGA-II-A and GTLP under specific power demands as shown in Figures 5(a) and 5(b), the performance comparison diagrams of DNSGA-II-A and GTLP under different power demands as shown in Figures 6(a), 6(b), 6(c) and 6(d), and Figure 7 the average emission improvement percentage data diagram obtained by comparing the emissions of the thermal power generating units of GTLP with those of DNSGA-II-A under different power demands as shown in Figure 5(a) and 5(b)It respectively shows the performance comparison between the two methods of DNSGA-II-A and GTLP under two power demands of 1000MW and 1300MW. The horizontal axis represents the total fuel cost of the thermal power generation units in the scheduling scheme, and the vertical axis represents the emissions of the thermal power generation units. Analyzing Figure 5(a), it can be seen that under the power demand of 1000MW, GTLP performs slightly worse in the -1.4% marked area, and has a significant improvement in the 4.2% marked area. However, overall, GTLP has an improvement of 3.6% compared to DNSGA-II-A, indicating that GTLP has better performance. Analyzing Figure 5(b), it can be seen that GTLP performs slightly worse in the -1.1% marked area, but has a significant improvement in the 4.7% marked area. Overall, GTLP shows an improvement of 3.5% compared to DNSGA-II-A. Figures 6(a), 6(b), 6(c) and 6(d) respectively show the performance comparison of minimizing the total fuel cost of the thermal power generation units and minimizing the emissions of the thermal power generation units between the two methods of DNSGA-II-A and GTLP under different power demands of 900MW, 1000MW, 1100MW and 1300MW. The horizontal axis represents the total fuel cost of the thermal power generation units in the scheduling scheme, and the vertical axis represents the emissions of the thermal power generation units. Analyzing Figure 6(a), it can be seen that GTLP has an average improvement percentage of 3.8% compared to DNSGA-II-A under the power demand of 900MW. Analyzing Figure 6(b), it can be seen that it has an average improvement percentage of 3.4% compared to DNSGA-II-A under the power demand of 1000MW. Analyzing Figure 6(c), it can be seen that it has an average improvement percentage of 3.5% compared to DNSGA-II-A under the power demand of 1100MW. Analyzing Figure 6(d), it can be seen that it has an average improvement percentage of 3.6% compared to DNSGA-II-A under the power demand of 1300MW. Figure 7 shows the average emission improvement percentage of the emissions of the thermal power generation units of GTLP compared to those of the thermal power generation units of DNSGA-II-A under different power demands. Analyzing Figure 7 it can be seen that GTLP has an average emission improvement of 3.9% compared to DNSGA-II-A under the power demand of 900MW, an average emission improvement of 3.6% under the power demand of 1000MW, an average emission improvement of 3.6% under the power demand of 1100MW, and an average emission improvement of 3.3% under the power demand of 1300MW. Thus, it can be obtained that the hydrothermal power generation system scheduling optimization method based on transfer learning and GAN of the present invention shows better stability compared to DNSGA-II-A under different power demands, and can continuously provide a good balance between minimizing the total fuel cost of the thermal power generation units and minimizing the emissions of the thermal power generation units.

[0081] In summary, the method for optimizing the scheduling of a hydrothermal power generation system based on transfer learning and GAN can, when the power demand changes, make full use of the population optimization experience of historical optimized scheduling schemes or extract the experience of the population of historical optimized scheduling schemes based on the idea of transfer learning, and combine the generative adversarial network GAN and the standard normal distribution to generate a new population of scheduling schemes with historical experience. Thus, it can maintain the population information of the scheduling scheme that minimizes the total fuel cost of thermal power generation units and minimizes the emissions of thermal power generation units while minimizing historical balance, adaptively adjust the information of the historical scheduling scheme population according to the relative change degree of the power demand, no longer strongly rely on artificial parameter settings, and the obtained initial scheduling scheme reduces the violent oscillation in the optimization process. During the continuous power demand change process, it can maintain the consistency and reliability of the scheduling schemes in the optimized scheduling scheme population under the current power demand obtained by optimization, and improve the stability in the effect of balancing the minimization of the total fuel cost of thermal power generation units and the minimization of the emissions of thermal power generation units.

Claims

1. A scheduling optimization method for a hydrothermal power generation system based on transfer learning and GAN. When the current power demand is the first power demand, a random normal distribution method is used to randomly generate multiple scheduling plans to form an initial scheduling plan population. Then, for each scheduling plan in the initial scheduling plan population, a non-dominated sorting genetic algorithm is used for selection, crossover, and mutation operations to obtain new scheduling plans to form an optimized scheduling plan population for the current power demand. A certain scheduling plan is selected from the optimized scheduling plan population for the current power demand as the scheduling plan for the current power demand to perform the scheduling of the hydrothermal power generation system; characterized in that: When the power demand changes and the current power demand after the change is the second power demand, by combining the randomly generated population of scheduling plans with the optimized population of scheduling plans for the previous power demand, an optimized population of scheduling plans for the current power demand is obtained based on the non-dominated sorting genetic algorithm. Then, a certain scheduling plan is selected from the optimized population of scheduling plans for the current power demand as the scheduling plan for the current power demand, and the hydro-thermal power generation system is scheduled; when the power demand changes and the current power demand after the change is the third power demand or a power demand after the third power demand, the experience of the historical optimized population of scheduling plans is extracted based on the idea of transfer learning, and a new population of scheduling plans with historical experience is generated by combining the generative adversarial network GAN and the standard normal distribution. An optimized population of scheduling plans for the current power demand is obtained through fusion and the non-dominated sorting genetic algorithm. Then, a certain scheduling plan is selected from the optimized population of scheduling plans under the current power demand as the scheduling plan for the current power demand, and the hydro-thermal power generation system is scheduled.

2. The hydrothermal power generation system scheduling optimization method based on transfer learning and GAN according to claim 1, characterized in that When the power demand changes and the current power demand after the change is the second power demand, the specific process of obtaining the optimized population of scheduling plans for the current power demand by combining the randomly generated population of scheduling plans with the optimized population of scheduling plans for the previous power demand based on the non-dominated sorting genetic algorithm is as follows: First, the random normal distribution method is used to randomly generate multiple scheduling plans to form a random population of scheduling plans. The currently obtained random population of scheduling plans and the optimized population of scheduling plans for the first power demand are mixed to form a temporary population of scheduling plans. Then, the required number of scheduling plans is selected from the temporary population of scheduling plans based on the non-dominated sorting operation to form the initial population of scheduling plans for the current power demand. Next, the selection, crossover, and mutation operations of the non-dominated sorting genetic algorithm are respectively performed on each scheduling plan in the initial population of scheduling plans for the current power demand to obtain new scheduling plans, which form the optimized population of scheduling plans for the current power demand.

3. The scheduling optimization method for a hydrothermal power generation system based on transfer learning and GAN according to claim 1, wherein When the power demand changes and the current power demand after the change is the third power demand or a power demand after the third power demand, the specific process of extracting the experience of the historical optimized population of scheduling plans based on the idea of transfer learning, and generating a new population of scheduling plans with historical experience by combining the generative adversarial network GAN and the standard normal distribution, and obtaining the optimized population of scheduling plans for the current power demand through fusion and the non-dominated sorting genetic algorithm is as follows: Step 1: Perform non-dominated sorting on the scheduling plans in the population of the optimization scheduling plan for the previous electricity demand, and divide all the scheduling plans in the population of the optimization scheduling plan for the previous electricity demand into different levels. After division, each scheduling plan in the first level is better than each scheduling plan in the second level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units. Each scheduling plan in the second level is better than each scheduling plan in the third level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units, and so on. That is, each scheduling plan in a lower level is better than each scheduling plan in a higher level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units; Step 2: Calculate the total fuel cost of the thermal power generation units and the emissions of the thermal power generation units for each scheduling plan in the first level. Use the total fuel cost of the thermal power generation units as the horizontal axis of a two-dimensional coordinate system and the emissions of the thermal power generation units as the vertical axis of the two-dimensional coordinate system to form a two-dimensional coordinate system. The total fuel cost of the thermal power generation units for each scheduling plan in the first level is used as its abscissa, and the emissions of the thermal power generation units are used as its ordinate to form its two-dimensional coordinate point. Mark the two-dimensional coordinate points of each scheduling plan in the first level in the two-dimensional coordinate system. Use the minimum value of the abscissas of the two-dimensional coordinate points of all the scheduling plans in the first level as the abscissa and the maximum value of the ordinates as the ordinate to form the first boundary coordinate, which is marked as boundary point one in the two-dimensional coordinate system. Use the maximum value of the abscissas among the two-dimensional coordinate points of all the scheduling plans in the first level as the abscissa and the minimum value of the ordinates as the ordinate to form the second boundary coordinate, which is marked as boundary point two in the two-dimensional coordinate system. Connect boundary point one and boundary point two with a straight line, and use the line segment between boundary point one and boundary point two as the extreme value line; Step 3: Segment the horizontal axis between the abscissa of boundary point one and the abscissa of boundary point two in the two-dimensional coordinate system to obtain multiple horizontal axis segments. The specific process is as follows: Set an integer greater than or equal to 1 as the number of segmentation segments. If the length value of each segment obtained by evenly segmenting this horizontal axis according to the number of segmentation segments is an integer or a decimal with no more than two digits after the decimal point, then evenly segment the horizontal axis between the abscissa of boundary point one and the abscissa of boundary point two according to the number of segmentation segments to obtain multiple horizontal axis segments. If the length value of each segment obtained is a decimal with more than two digits after the decimal point, then retain two digits after the decimal point for this length value using the truncation method to obtain the truncated segment length value. Starting from the abscissa of boundary point one, segment this horizontal axis with the truncated segment length value until the remaining length of a horizontal axis segment is less than the truncated segment length value, and use it as the last horizontal axis segment to obtain multiple horizontal axis segments; Step 4: Straight lines perpendicular to the horizontal axis of the two-dimensional coordinate system are drawn at the two endpoints of each horizontal axis segment. The two straight lines perpendicular to the horizontal axis of the two-dimensional coordinate system drawn at the two endpoints of each horizontal axis segment are called its two perpendicular lines. The two perpendicular lines of each horizontal axis segment will intersect with the extreme value line. A quadrilateral region is formed among the two perpendicular lines of each horizontal axis segment, the extreme value line, and the horizontal axis of the two-dimensional coordinate system, obtaining multiple quadrilateral regions. At this time, the two-dimensional coordinate points of all scheduling schemes at the first level are divided into these quadrilateral regions; Step 5: The direction from the abscissa of boundary point 1 to the abscissa of boundary point 2 is taken as the left-to-right direction; if the two-dimensional coordinate point of a certain scheduling scheme at the first level falls on the perpendicular line drawn at the left endpoint of the leftmost horizontal axis segment, it is considered to be in the leftmost quadrilateral region. If the two-dimensional coordinate point of a certain scheduling scheme at the first level falls on the perpendicular line drawn at the right endpoint of the rightmost horizontal axis segment, it is considered to be in the rightmost quadrilateral region. If the two-dimensional coordinate point of a certain scheduling scheme falls on other perpendicular lines except the perpendicular line drawn at the left endpoint of the leftmost horizontal axis segment and the perpendicular line drawn at the right endpoint of the rightmost horizontal axis segment, it is considered to be in the quadrilateral region with this perpendicular line as the left side; Step 6: Calculate the perpendicular line distance from the two-dimensional coordinate point of each scheduling scheme at the first level to the extreme value line. Select the two-dimensional coordinate point with the largest perpendicular line distance from all two-dimensional coordinate points in each quadrilateral region to the extreme value line as an alternative two-dimensional coordinate point. If there are multiple, randomly select one of them. The scheduling schemes corresponding to all the alternative two-dimensional coordinate points obtained at this time form the first scheduling scheme source population; Step 7: Calculate the mutual information values between each scheduling scheme in the optimized scheduling scheme population of the previous power demand of the current power demand and each scheduling scheme in the optimized scheduling scheme population of the pre-previous power demand of the current power demand. All the calculated mutual information values form a mutual information matrix. Each row of the mutual information matrix represents the mutual information values between a scheduling scheme in the optimized scheduling scheme of the previous power demand and all the scheduling schemes in the optimized scheduling scheme population of the pre-previous power demand of the current power demand. Each element in each column of each row of the mutual information matrix represents the mutual information value between a scheduling scheme in the optimized scheduling scheme of the previous power demand and a scheduling scheme in the optimized scheduling scheme population of the pre-previous power demand of the current power demand. Each column of the mutual information matrix represents the mutual information values between each scheduling scheme in the optimized scheduling scheme of the previous power demand and a certain scheduling scheme in the optimized scheduling scheme population of the pre-previous power demand of the current power demand. Each element in each row of each column of the mutual information matrix represents the mutual information value between a scheduling scheme in the optimized scheduling scheme of the pre-previous power demand and a scheduling scheme in the optimized scheduling scheme population of the previous power demand of the current power demand; Step 8: Use the mutual information value between a certain scheduling plan in the population of optimized scheduling plans for the previous power demand of the current power demand and a certain scheduling plan in the population of optimized scheduling plans for the power demand before the previous one of the current power demand as the preference value between the two. For each scheduling plan in the population of optimized scheduling plans for the previous power demand of the current power demand, sort the mutual information values between it and all scheduling plans in the population of optimized scheduling plans for the power demand before the previous one of the current power demand from high to low to form its preference list, so as to obtain the preference list of each scheduling plan in the population of optimized scheduling plans for the previous power demand of the current power demand. For each scheduling plan in the population of optimized scheduling plans for the power demand before the previous one of the current power demand, sort the mutual information values between it and all scheduling plans in the population of optimized scheduling plans for the previous power demand of the current power demand from high to low to form its preference list, so as to obtain the preference list of each scheduling plan in the population of optimized scheduling plans for the power demand before the previous one of the current power demand; Step 9: According to the preference list of each scheduling plan in the population of optimized scheduling plans for the previous power demand of the current power demand and the preference list of each scheduling plan in the population of optimized scheduling plans for the power demand before the previous one of the current power demand, adopt the stable marriage matching strategy to match between each scheduling plan in the population of optimized scheduling plans for the power demand before the previous one and each scheduling plan in the population of optimized scheduling plans for the previous power demand, so that the scheduling plans in the population of optimized scheduling plans for the previous power demand of the current power demand are in one-to-one correspondence with the scheduling plans in the population of optimized scheduling plans for the power demand before the previous one of the current power demand, and obtain multiple stable matching scheduling plan pairs. Each stable matching scheduling plan pair is composed of a scheduling plan in the population of optimized scheduling plans for the previous power demand of the current power demand and a scheduling plan in the population of optimized scheduling plans for the power demand before the previous one of the current power demand; Step 10: Perform non-dominated sorting on the population of optimized scheduling schemes for the previous power demand of the current power demand, and divide all the scheduling schemes in the population of optimized scheduling schemes for the previous power demand into different levels. After division, each scheduling scheme in the first level is better than each scheduling scheme in the second level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units. Each scheduling scheme in the second level is better than each scheduling scheme in the third level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units, and so on. That is, each scheduling scheme in a lower level is better than each scheduling scheme in a higher level in terms of minimizing the total combustion cost of thermal power generation units and minimizing the emissions of thermal power generation units. Sort the scheduling schemes in each level from the first level to the last level to obtain the sorting of all the scheduling schemes in the population of optimized scheduling schemes for the previous power demand of the current power demand. According to the sorting of all the scheduling schemes in the population of optimized scheduling schemes for the previous power demand of the current power demand, sort the obtained multiple stable matching scheduling scheme pairs to obtain the sorted stable matching scheduling scheme pairs, and then select the first 50% of the stable matching scheduling scheme pairs from the sorted stable matching scheduling scheme pairs. If 50% of the total number of stable matching scheduling scheme pairs is not an integer, round up. Use the scheduling schemes in all the selected stable matching scheduling scheme pairs to form the scheduling scheme source population two. Step 11: Merge the scheduling scheme source population one and the scheduling scheme source population two to form the scheduling scheme source population three. Step 12: Calculate the mean of the scheduling scheme source population one, the mean of the scheduling scheme source population two, the total combustion cost of the thermal power generation units and the emissions of the thermal power generation units of each scheduling scheme in the scheduling scheme source population one for the current power demand, and the total combustion cost of the thermal power generation units and the emissions of the thermal power generation units of each scheduling scheme in the scheduling scheme source population one for the previous power demand of the current power demand respectively. Obtain the relative change degree between the current power demand and the previous power demand according to the relative change amount between the total combustion cost of the thermal power generation units and the emissions of the thermal power generation units of each scheduling scheme in the scheduling scheme source population one for the current power demand and the total combustion cost of the thermal power generation units and the emissions of the thermal power generation units of the previous power demand. The value range of the relative change degree is [0, 1]. Use the relative change degree to perform weighted summation on the mean of the scheduling scheme source population one and the mean of the scheduling scheme source population two respectively, and use the weighted summation result obtained at this time as the adaptive comprehensive mean of the scheduling scheme source population three. Then calculate the covariance of the scheduling scheme source three. Step 13: Set the input of the generator of the generative adversarial network (GAN) to a multivariate Gaussian distribution that conforms to the adaptive comprehensive mean and covariance of the source population three of the scheduling scheme. Use each scheduling scheme in the optimized scheduling scheme population of the previous electricity demand as the input of the discriminator of the GAN. Conduct adversarial training on the generator and discriminator of the GAN to obtain the trained GAN, and use the trained GAN to generate the source population four of the scheduling scheme. Step 14: Perform Cholesky decomposition on the covariance matrix of the source population three of the scheduling scheme to obtain a lower triangular matrix. Randomly generate multiple scheduling schemes using the standard normal distribution method. Multiply each generated scheduling scheme by the lower triangular matrix respectively to obtain a scheduling scheme with a specified covariance matrix corresponding to each scheduling scheme. Add the adaptive comprehensive mean of the source population three of the scheduling scheme to each of the currently obtained scheduling schemes with a specified covariance matrix to obtain new scheduling schemes, and use these new scheduling schemes to form the source population five of the scheduling scheme. Step 15: Merge the source population four of the scheduling scheme, the source population five of the scheduling scheme, and the optimized scheduling scheme population of the previous electricity demand to form the source population six of the scheduling scheme. Step 16: Use non-dominated sorting operation to select the required number of scheduling schemes from the source population six of the scheduling scheme to form the initial scheduling scheme population under the current electricity demand. Then, use the non-dominated sorting genetic algorithm to perform selection, crossover, and mutation operations on the initial scheduling scheme population under the current electricity demand to obtain new scheduling schemes to form the optimized scheduling scheme population of the current electricity demand.

4. The scheduling optimization method for a hydrothermal power generation system based on transfer learning and GAN according to claim 3, wherein The means of the source population one of the scheduling scheme and the source population two of the scheduling scheme are calculated using formulas (1) and (2) respectively: Among them, μ k is the mean of the first source population of scheduling schemes, S is the total number of scheduling schemes of the first source population of scheduling schemes, and kind i is the i-th scheduling scheme in the first source population of scheduling schemes, where i = 1, 2, …, S, and μ r is the mean of the second source population of scheduling schemes, Z is the medium number of scheduling schemes of the second source population of scheduling schemes, and rind j is the j-th scheduling scheme in the second source population of scheduling schemes, where j = 1, 2, …, Z.

5. The scheduling optimization method for a hydrothermal power generation system based on transfer learning and GAN according to claim 4, characterized in that The relative change degree between the current electricity demand and the previous electricity demand of the current electricity demand is calculated using formula (3): Among them, M is the number of objective functions in the mathematical model obtained after digital-analog modeling of the hydrothermal power generation scheduling system, v = 1, 2, …, M, u = 1, 2, …, S, f v (x u ) is the function value of the v-th objective function of the u-th scheduling plan x u in the first population of scheduling plan sources obtained through the mathematical model under the current power demand, is the function value of the v-th objective function of the u-th scheduling plan x u in the previous power demand of the current power demand obtained through the mathematical model, Max v is the maximum function value among the function values of the v-th objective function of all scheduling plans in the first population of scheduling plan sources obtained through the mathematical model, Min v is the minimum function value among the function values of the v-th objective function of all scheduling plans in the first population of scheduling plan sources obtained through the mathematical model, δ is the relative change degree in the range of [0, 1] between the current power demand and the previous power demand of the current power demand, and ︱︱ is the absolute value symbol.

6. The scheduling optimization method for a hydrothermal power generation system based on transfer learning and GAN according to claim 5, characterized in that The adaptive comprehensive mean of the source population three of the scheduling scheme is calculated using formula (4): MV = δμ k +(1 - δ)μ r (4) where MV is the adaptive comprehensive mean of the source population three of the scheduling scheme.

7. The scheduling optimization method for a hydrothermal power generation system based on transfer learning and GAN according to claim 3, characterized in that The generator of the generative adversarial network (GAN) adopts a multi-layer feedforward neural network. The generator contains two hidden layers and an output layer. The first hidden layer of the generator is configured with 512 neurons, and the second hidden layer of the generator is configured with 256 neurons. The first hidden layer of the generator uses the LeakyReLU activation function, and the second hidden layer of the generator uses the ELU activation function. The output layer of the generator outputs the scheduling scheme generated by the generator through the Tanh activation function. The input of the generator is the scheduling scheme sampled from the multivariate Gaussian distribution that conforms to the adaptive comprehensive mean and covariance of the scheduling scheme source population three. The discriminator of the generative adversarial network (GAN) adopts a lightweight fully connected network structure. The discriminator of the generative adversarial network (GAN) contains a hidden layer and an output layer. The hidden layer of the discriminator is configured with 256 neurons and uses Swish as the activation function. The input of the discriminator is each scheduling scheme in the optimized scheduling scheme population of the previous power demand and each scheduling scheme generated by the generator. Each scheduling scheme in the optimized scheduling scheme population of the previous power demand serves as the real dataset of the discriminator, and each scheduling scheme generated by the generator serves as the generated dataset of the discriminator. The output layer of the discriminator uses the Sigmoid activation function to obtain the probability that each scheduling scheme in the real dataset and the generated dataset is a real scheduling scheme. When performing adversarial training on the generator and discriminator of the generative adversarial network (GAN), the batch size is set to 32, the number of training epochs is set to 100, and the gradient descent method is used to update the generator and discriminator of the GAN during the adversarial training process.

Images