A deep learning method for solving unit commitment problems considering confidence intervals
By using deep learning networks to optimize the solution process of unit combination problems in power systems, combined with confidence interval processing and mathematical planning model, the problem of too long solving in traditional methods is solved, and faster solution speed and result accuracy are achieved.
Patent Information
- Application Number
- CN202411230246.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-04
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-09-04
AI Technical Summary
Due to its non-determinism and computational complexity in modern power systems, traditional solutions are difficult to improve the solution speed while ensuring accuracy, especially when fast response is required.
A method for solving the problem of combining deep learning units considering confidence intervals is proposed. By scene generation of load and wind power data, deep learning network is trained, output results are optimized, and processed state variables are input into the mathematical planning model for solving, which significantly reduces Solve time.
On the premise of ensuring the quality of the result, the solution time of unit combination problems is significantly reduced, the solution speed is improved, and the fast response needs are met.
Smart Images

Figure CN118735309B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power system planning, and in particular to a method for solving a deep learning unit commitment problem considering confidence intervals. Background Art
[0002] In modern power systems, unit commitment (UC) plays a vital role in ensuring the stable supply and effective distribution of electricity. When a large number of new energy units are connected to the power grid, the uncertainty and volatility of their output pose a huge challenge to the stability of the power grid.
[0003] The unit commitment problem is essentially a non-deterministic polynomial (NP)-hard problem, which means it is computationally extremely challenging. With the continuous expansion of the scale of power systems, the number of decision variables and constraints in the model has increased dramatically, making it more difficult to solve the unit commitment problem, especially when rapid response is required. Traditional operations research solution ideas (such as establishing mixed integer linear programming models for solution) are difficult to meet actual needs in terms of solution efficiency. Therefore, how to improve the solution speed while ensuring the solution accuracy has become the focus of current research.
[0004] In recent years, with the development of artificial intelligence technology, using machine learning (ML) as a representative method to solve unit commitment problems is becoming a new research idea. Machine learning methods can discover potential laws and patterns by learning from historical data, thereby providing near-optimal solutions in a short period of time. Summary of the invention
[0005] In view of the fact that it takes too long to construct a mathematical programming model to solve the unit commitment problem, the present invention proposes a deep learning unit commitment problem solving method considering confidence intervals. By optimizing and fixing the deep learning output results, not only the result quality is guaranteed, but also the solution time is significantly reduced.
[0006] The method proposed in the present invention is mainly divided into two stages:
[0007] In the first stage, scenarios are generated for load and wind power data to obtain a data set containing load scenarios, wind power scenarios, and unit state variables, which are then input into a deep learning network for training to obtain the initial unit state variable results.
[0008] In the second stage, different confidence intervals are divided according to the errors of the unit state variables, and the unit state variables in each interval are processed separately. Finally, the processed state variables are input into the mathematical programming model for solution to obtain the unit power and state variable results that rigidly meet the physical constraints. BRIEF DESCRIPTION OF THE DRAWINGS
[0009] Figure 1 A flow chart of a method for solving the unit commitment problem with deep learning considering confidence intervals;
[0010] Figure 2 Deep learning network model structure diagram;
[0011] Figure 3 Decision diagram for treatments at different confidence intervals. DETAILED DESCRIPTION
[0012] First, a set of benchmark data is obtained by solving the unit combination mathematical programming model ,in, , , .
[0013] Secondly, the benchmark data To generate a scene, use the benchmark data The value in itself is taken as the expected value of the normal distribution, and the standard deviation is 10% of its own value, generate samples through Monte Carlo simulation and Latin hypercube sampling, and randomly select m samples from the generated samples and compare them with the benchmark data Together they form the input of a deep learning network model. Specifically, the data sample content of a deep learning network model is .in , .
[0014] In the formula Wind power A collection of The total set ; Load power A collection of The total set ; is the set of times t; is the set of units g; is the state variable of unit g at time t (a binary variable, 1 represents that unit g is in the on state at time t, and 0 represents that unit g is in the off state at time t)
[0015] The deep learning network used in the present invention is composed of multiple long short-term memory (LSTM) networks, a dropout layer, and a dense layer. The input of the deep learning network is a multi-scenario data set generated based on benchmark wind power and load power. , the output is the state variable of the unit in each time period The activation function selected in the Dense layer is the sigmoid function.
[0016] Next is the division of confidence intervals. First, the unit status results output by the deep learning network model are calculated. The unit status results calculated by the mathematical programming model Then the absolute difference is calculated by summing up the absolute differences to calculate the error between the unit status result output by the deep learning network model and the unit status result calculated by the mathematical programming model. Then, according to the calculated error The unit status results output by the deep learning network model are divided into different confidence intervals. Different optimization strategies are adopted according to different confidence intervals.
[0017]
[0018]
[0019] Where: for and The absolute difference between them, N represents the number of samples, and T is the number of time periods.
[0020] When the output results of the deep learning network of unit g at all times in N samples are completely consistent with the results calculated by the mathematical programming model, that is, , indicating that the prediction accuracy of the on / off state of unit g has reached 100%. In this case, since the on / off state of unit g always meets the minimum start / stop time and logical constraints, no additional feasibility check is required. When the results of this interval are brought into the mathematical programming model for solution, these variables will be directly set to fixed values.
[0021] For the unit start-stop decision with error in this confidence interval, that is, We believe that the state results output by the deep learning network are of high quality. However, due to slight errors, the feasibility of this part of the deep learning output results still needs to be checked to ensure that it meets the constraints of the unit mathematical programming model.
[0022] The objective function of the feasibility check is:
[0023]
[0024] The constraints for the feasibility check are:
[0025]
[0026]
[0027]
[0028]
[0029]
[0030] Where: Yes The feasible solution with the smallest Euclidean distance; is the startup status of unit g at time t (binary variable, 1 means the unit is started, 0 means the unit is not started); is the shutdown status of unit g at time t (binary variable, 1 means the unit is shut down, 0 means the unit is not shut down); is the minimum start-up time of unit g; is the minimum shutdown time of unit g; The initial state of unit g (binary variable, 1 means that unit g is in the on state at the initial moment; 0 means that unit g is in the off state at the initial moment); is the time that the unit g must remain in operation before dispatch begins; is the time that the unit g must remain shut down before dispatch begins; is the time that unit g has been on before dispatch; is the time that unit g has remained shut down before being dispatched.
[0031] After the feasibility check, we use the newly generated start and stop decisions Replace the initial output Since the prediction error in this interval is extremely small, we can assume that the start-stop decision obtained after the feasibility check is basically consistent with the optimal solution. Therefore, when the start-stop decision in this interval is brought into the mathematical programming model for subsequent solution, the output result of this interval is also set to a fixed value.
[0032] For the error in The prediction error of the start and stop decision of the unit in this interval cannot be ignored. Even after the feasibility check, the result is still some distance away from the optimal solution. It cannot be set as a fixed value in the subsequent solution. However, this part of the results still has reference value for the subsequent solution, so we will not discard it. We will set this part of the results as the hot start initial value of the model in the subsequent solution to provide a better initial solution.
[0033] For the error in The start-stop decision in . We believe that this part of the results has a large error compared with the actual results and has little reference value for subsequent solutions, so we directly discard it. In the subsequent mathematical programming model solution process, this part of the results will be directly set as the system default value.
[0034] After processing the unit status of each interval, we bring the obtained unit status into the mathematical programming model for solution; the unit combination mathematical programming model is as follows:
[0035] The objective function of the model includes startup type cost, startup phase cost and production cost:
[0036]
[0037]
[0038]
[0039]
[0040] is the startup type cost of unit g at time t; is the startup cost of unit g at time t; is the production cost of unit g at time t; is the hot start cost of unit g; is the cold start cost of unit g; is the time until time t when unit g has been shut down; is the time required for the unit to start cold; is the cost coefficient of unit g; is the power size of the i-th stage during the startup process of unit g; is the index of the unit start method; It is the starting process of unit g; is the startup type s of unit g at time t (a binary variable, 1 means that unit g adopts the startup method of type s; 0 means that unit g does not adopt the startup method of type s); is the actual power of unit g at time t;
[0041] Load constraints of thermal power units:
[0042]
[0043] Where: is the wind power at time t; is the load size at time t.
[0044] Power constraints of thermal power units:
[0045]
[0046]
[0047]
[0048] Where: is the maximum output power of unit g; is the minimum output power of unit g; is the power level of unit g at time t when it is above the minimum output power of the unit; is the power size of the i-th stage during the shutdown process of unit g; It is the shutdown process of the unit.
[0049] Thermal power unit climbing constraints:
[0050]
[0051]
[0052] Where: is the upper spinning reserve capacity of unit g at time t; is the lower spinning reserve capacity of unit g at time t; is the gradeability of the unit g; is the downhill capability of the unit g;
[0053] The climbing constraint for the first moment is considered separately:
[0054]
[0055]
[0056]
[0057] Where: It is the power of unit g at the initial moment (i.e. the last moment of the previous day).
[0058] Spinning reserve constraints for thermal power units:
[0059]
[0060]
[0061] Where: is the upper spinning reserve capacity required at time t; is the lower spinning reserve capacity required at time t.
[0062] Minimum start / shutdown time constraints for thermal power units:
[0063]
[0064]
[0065]
[0066]
[0067]
[0068] Thermal power unit startup / shutdown logic constraints:
[0069]
[0070]
[0071] Selection of starting type for thermal power units:
[0072]
[0073]
[0074] In order to clearly illustrate the advantages of the method proposed in this patent, Table 1 compares the time for solving the unit combination problem by the method of this patent with the time for solving the unit combination problem by a pure mathematical programming model.
[0075] Table 1 Comparison of model solution time
[0076] Test system 10 units 32 units 69 units Solution 1 (seconds) 1212.3 3069.3 6836.1 Solution 2 (seconds) 4501.8 10106.5 11909.6
[0077] As shown in Table 1, Scheme 1 is the time required for a deep learning unit combination problem solving method considering confidence intervals proposed in this patent. Scheme 2 is the solution time of a pure mathematical programming model. The results show that Scheme 1 has significantly improved the solution time.
[0078] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or exchanges made by technicians in the technical field on the basis of the present invention are all within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.
Claims
1. A method for solving a deep learning unit commitment problem considering confidence intervals, characterized in that: The method comprises the following operations: In the first stage, a set of baseline wind power and load data is obtained. Secondly, use To generate a scene, use the benchmark data The value in itself is taken as the expected value of the normal distribution, and the standard deviation is 10% of its own value, generate samples through Monte Carlo simulation and Latin hypercube sampling, and select m samples from the generated samples to compare with the benchmark data Together they form the input of the deep learning network model; In the second stage, different confidence intervals are divided according to the unit start-stop decision error, and the unit decisions in each interval are processed separately: First, the unit start-stop decision error α is calculated. g , the formula is as follows: Where: e g,t for and The absolute difference between It is the unit status result output by the deep learning network model; is the unit status result calculated by the mathematical programming model; N represents the number of samples; T is the number of time periods; The second is the division and processing of confidence intervals: When α g =0, indicating that the prediction accuracy of the on / off state of unit g has reached 100%; in this case, since the start / stop decision of unit g always meets the minimum start / stop time and logical constraints, no additional feasibility check is required; when the results of this interval are brought into the mathematical programming model for solution, these variables will be directly set to fixed values; When 0<α g When ≤R1, the start-stop decision quality of deep learning output is high; however, due to the existence of errors, it is still necessary to perform a feasibility check on this part of the deep learning output results to ensure that it meets the constraints of the unit mathematical programming model; the feasibility check is as follows: Where: Yes The feasible solution with the smallest Euclidean distance; v g,i is the startup status of unit g at time i, a binary variable, 1 indicates that the unit is started, and 0 indicates that the unit is not started; w g,i is the shutdown status of unit g at time i, a binary variable, 1 means the unit is shut down, 0 means the unit is not shut down; is the minimum start-up time of unit g; is the minimum shutdown time of unit g; The initial state of unit g, a binary variable, 1 means that unit g is in the on state at the initial moment; 0 means that unit g is in the off state at the initial moment; TU g is the time that unit g must remain in operation before scheduling begins; TD g is the time that the unit g must remain shut down before dispatch begins; is the time that unit g has been on before dispatch; is the time that unit g has remained shut down before dispatch; After the feasibility check, use the newly generated start and stop decision Replace the initial output Since the prediction error in this interval is extremely small, the start-stop decision obtained after the feasibility check is consistent with the optimal solution; therefore, when the start-stop decision in this interval is brought into the mathematical programming model for subsequent solution, the output result of this interval is also set to a fixed value; When R1<α g When ≤R2, since the prediction error cannot be ignored, even after the feasibility check, the result is still some distance away from the optimal solution; Therefore, the results of this part It cannot be set as a fixed value in the subsequent solution; however, this part of the results still has reference value for the subsequent solution, so it will not be discarded; instead, this part of the results will be set as the hot start initial value of the model in the subsequent solution to provide a better initial solution; When α g When the error is greater than R2, the results are discarded directly because they are not very valuable for subsequent solutions. In the subsequent mathematical programming model solving process, the results of this part will be directly set as the system default value; Finally, the processed start-stop decision is input into the mathematical programming model for solution to obtain the unit power and start-stop decision scheme that rigidly meets the physical constraints.
