A method for extracting reservoir operation rules based on rolling simulation strategy

Abstract

This invention provides a method for extracting reservoir scheduling rules based on a rolling simulation strategy, comprising the following steps: given an initial water level Z0 during the reservoir scheduling period, where the scheduling time period t = 0; using the initial water level Z0... t and inbound flow I t As a characteristic factor, with the final water level Z t+1 For the target value, train model M t M t This represents the scheduling rules for the reservoir during the t-th scheduling period; the initial water level Z... t and inbound flow I t Input model M t The decision level K is obtained. t+1 ; Determine the decision level K t+1 Does it violate reservoir scheduling constraints? If so, correct it; set the decision water level K. t+1 Set the initial water level Z for the next scheduling period t+1 Then, set t = t + 1; determine whether the scheduling period t meets the termination condition. If so, output the final scheduling rule; otherwise, repeat steps 2 to 6. This improves the practicality of the scheduling rule and the power generation efficiency of the reservoir system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of reservoir operation, and particularly relates to a reservoir operation rule extraction method based on a rolling simulation strategy. BACKGROUND

[0002] Reservoir operation is to maximize the economic and social benefits of the reservoir system under the premise of ensuring flood control safety by reasonably arranging the reservoir's water storage and release operation. As the core tool for guiding the operation of the reservoir system, the reservoir operation rule not only affects the planning and construction of water conservancy facilities, but also is a key factor to ensure the comprehensive benefits during the operation and management period of the reservoir. With the development of social economy and the influence of environmental changes, reservoir operation is facing more complex challenges, such as the increase of hydrological uncertainty caused by climate change and the improvement of ecological environment protection demand.

[0003] Traditional reservoir operation methods mostly rely on empirical formulas and expert knowledge. Although this method can meet the needs of reservoir operation to some extent, it is difficult to adapt to complex and variable hydrological conditions due to the lack of in-depth analysis of historical data and consideration of future uncertainty. In recent years, with the development of information technology, especially data mining technology, data-driven methods have been widely used in the extraction of reservoir operation rules. These methods include but are not limited to artificial neural networks (ANN), support vector machines (SVM), genetic algorithms (GA), etc. These advanced data processing technologies can mine potential patterns and rules from a large amount of historical data, providing a new way for formulating more scientific and reasonable operation strategies.

[0004] Although the data-driven reservoir operation rule extraction method has made some progress, there are still some deficiencies. For example, most existing methods often ignore the problem of error accumulation in the decision-making process when building the operation model, which may cause the model to gradually deviate from the actual operation state in the long-term operation, thereby affecting the reliability and accuracy of the operation rule. In addition, the existing methods have limited ability to respond to extreme weather events. When encountering rare high or low flow conditions, the performance of the model will be greatly reduced. These problems limit the effectiveness of existing methods in practical applications, and there is an urgent need to find new solutions. SUMMARY

[0005] In view of the deficiencies in the prior art, the present application provides a reservoir operation rule extraction method based on a rolling simulation strategy. In order to achieve the above purpose, the present application adopts the following technical solutions:

[0006] A reservoir operation rule extraction method based on a rolling simulation strategy, comprising the following steps:

[0007] Step 1, given the initial water level Z0 of the reservoir scheduling period, at this time, the scheduling period t = 0;

[0008] Step 2, with the initial water level Z t and the inflow I t as characteristic factors, the final water level Z t+1 as the target value, the model M t is trained t , wherein M t represents the scheduling rule of the reservoir in the tth scheduling period;

[0009] Step 3, input the initial water level Z t and the inflow I t into the model M t , and obtain the decision water level K t+1 ;

[0010] Step 4, judge whether the decision water level K t+1 violates the reservoir scheduling constraint, if it violates the constraint, correct it;

[0011] Step 5, set the decision water level K t+1 as the initial water level Z t+1 of the next scheduling period, and let t = t + 1;

[0012] Step 6: judge whether the scheduling period t meets the end condition, if yes, output the final scheduling rule, otherwise, execute steps 2-6 cyclically.

[0013] Further, in step 3, the decision water level K t+1 is the final water level predicted by the model M t , and there is a certain deviation value between the final water level Z t+1 obtained by the optimized scheduling.

[0014] Further, in step 4, the specific method of correction is that if the decision water level exceeds the upper limit value, it is forced to be equal to the upper limit value, and if the decision water level exceeds the lower limit value, it is forced to be equal to the lower limit value.

[0015] Further, in steps 2 and 3, the model includes a long short-term memory neural network model, an artificial neural network model, a light gradient boosting machine model, and a linear regression model.

[0016] Compared with the prior art, the present application has the following beneficial effects:

[0017] 1. The present application proposes a rolling simulation strategy with decision error correction capability. By introducing a decision error correction mechanism, the method can correct the model prediction results after each scheduling period, effectively reducing the accumulation of errors in the decision-making process. This improvement makes the scheduling rules more close to the actual operating conditions, improves the robustness and adaptability of the model, and even in the face of uncertain future hydrological conditions, it can maintain high scheduling accuracy.

[0018] 2. By using advanced machine learning algorithms such as long short-term memory networks (LSTM) and combining rolling simulation strategies, the present application can automatically learn and extract effective scheduling rules from a large amount of historical data. Compared with traditional methods, the present application not only simplifies the rule extraction process, but also improves the practicality of the rules. Specifically, the rules extracted by the present application can better balance the relationship between power generation benefits and environmental protection, providing technical support for sustainable development.

[0019] 3. By reducing the impact of decision error accumulation, the scheduling rules extracted by the present application can more accurately reflect the operating characteristics of the reservoir system, thereby helping to improve the power generation efficiency of the reservoir system. Experimental results show that the LSTM-RS model proposed by the present application performs excellently in terms of power generation, with only a difference of 0.08 million kilowatt-hours compared to the optimization scheme, and an increase of 0.19%, 0.95% and 0.95% in power generation compared to other models. This indicates that the present application not only has theoretical innovation, but also shows significant advantages in practical application.

[0020] 4. In addition to economic benefits, the present application also considers the needs of ecological and environmental protection. By accurately controlling the water level of the reservoir, it avoids the ecological and environmental damage caused by excessive water release or storage, and helps to maintain the health and stability of the river ecosystem. This is of great significance for promoting the sustainable use of water resources and achieving a win-win situation between economic and social development and environmental protection. BRIEF DESCRIPTION OF DRAWINGS

[0021] The present application will be further described below in conjunction with the drawings and examples:

[0022] Figure 1 A certain river basin map for the present application example;

[0023] Figure 2 Optimized scheduling results of a certain reservoir from 1959 to 2022 for the present application example;

[0024] Figure 3 Comparison of rolling simulation strategy effects for the present application example;

[0025] Figure 4 Comparison of water level simulation processes under each rule for the present application example. DETAILED DESCRIPTION

[0026] The technical solutions in the present application will be further described below with reference to the accompanying drawings and embodiments.

[0027] Please refer to Figures 1 to 4 A reservoir scheduling rule extraction method based on a rolling simulation strategy, comprising the following steps:

[0028] Step 1, the initial water level Z0 of the reservoir scheduling period is given, and the scheduling period t = 0 at this time;

[0029] Step 2, the initial water level Z t and the inflow I t are taken as characteristic factors, the final water level Z t+1 is taken as the target value, and the model M t is trained, wherein M t represents the scheduling rule of the reservoir in the tth scheduling period;

[0030] Step 3, the initial water level Z t and the inflow I t are input into the model M t , and the decision water level K t+1 is obtained; the decision water level Kt+1 is the final water level predicted by the model Mt, and there is a certain deviation value between the final water level Zt+1 obtained by the optimized scheduling;

[0031] Step 4, whether the decision water level K t+1 violates the reservoir scheduling constraint is judged, and if it violates the constraint, it is corrected; the specific method of correction is that if the decision water level exceeds the upper limit value, it is forced to be equal to the upper limit value, and if the decision water level exceeds the lower limit value, it is forced to be equal to the lower limit value;

[0032] Step 5, the decision water level K t+1 is set as the initial water level Z t+1 of the next scheduling period, and t = t + 1;

[0033] Step 6: whether the scheduling period t meets the end condition is judged, if yes, the final scheduling rule is output, otherwise, steps 2-6 are executed in a loop.

[0034] In order to verify the effectiveness of the proposed method, the research object and data information are presented:

[0035] The study object is introduced as follows: A river is 441.9 km long, with a drop of 1932 m, an average sediment concentration of 0.437%, and a total basin area of 26600 km2. There are 13 tributaries in the basin. The basin belongs to a typical subtropical plateau monsoon climate, with abundant rainfall, but unevenly distributed throughout the year, mainly concentrated in the flood season from May to September, with little interannual variation. The basin has abundant water resources and hydropower resources, with a dry river water level drop, good topographic and geological conditions, and a well-regulated cascade reservoir group.

[0036] The A station, B station, C station and D station sub-cascade hydropower stations on the main stream of the river are the key forces of upstream hydropower production. Among the four cascade reservoirs, A station, C station and D station are daily regulation reservoirs without long-term dispatching rule extraction conditions, while B station is a multi-year regulation reservoir with long-term dispatching rule extraction conditions. Therefore, this study mainly focuses on the extraction of B station reservoir dispatching rules. Figure 3 The spatial distribution of the four cascade reservoirs in the river basin is shown. The main characteristics of the B station reservoir are shown in Table 1.

[0037] Table 1 Design parameters of Guangzhao Reservoir

[0038]

[0039] Data information is introduced as follows: In this study, the time step of deterministic optimization scheduling and dispatching rule extraction is monthly. First, based on the long sequence of inflow data of the four cascade reservoirs of the river from 1959 to 2022, a total of 64 years, the differential evolution algorithm is used to obtain the deterministic optimization scheduling results. In the optimization scheduling model, the initial and final water levels of B station reservoir are set to 735 m, the output coefficient is set to 9.14, the minimum output is set to 180.2 MW, the maximum output is set to the expected maximum output under a specific water head, and the maximum discharge is set to the reservoir discharge capacity under a specific water head. The optimization scheduling results of B station reservoir from 1959 to 2022 are shown in Figure 2

[0040] After that, the optimization scheduling results are used to construct the dispatching rule extraction sample, i.e. the optimization scheduling data is divided into two subsets, including the training set and the test set. The training set is the optimization scheduling data from 1959 to 2016, a total of 58 years, which is used to train the model parameters; the test set is the optimization scheduling data from 2017 to 2022, a total of 6 years, which is used to test the model performance.

[0041] ​On the basis of the above objects, eight models of extraction operation rules are proposed, including Long Short-Term Memory (LSTM), Artificial Neural Network (ANN), Light Gradient Boosting Machine (LGB), Linear Regression (LR), LSTM-RS, ANN-RS, LGB-RS and LR-RS. Among them, LSTM-RS refers to the LSTM model using a rolling simulation strategy, and ANN-RS, LGB-RS and LR-RS use similar naming methods. In order to obtain better model performance, Bayesian optimization algorithm is used for hyperparameter optimization of all models. The models used in this study and their related parameters are shown in Table 2. In the scheduling rule extraction experiment, all models use unified input and output, the input (feature factor) is the inflow I t and the initial water level Z t , and the output (decision variable) is the final water level Z t+1 .

[0042] Table 2 Models and their parameters used in the study

[0043]

[0044] Results analysis: This study uses two indicators, power generation and decision water level deviation, to evaluate the pros and cons of the rules extracted by different models. Among them, the decision water level deviation is the root mean square error between the decision water level and the optimal water level. Power generation is used to measure the power generation benefit achieved by the operation rules. The greater the power generation, the better the operation rules, and vice versa. The decision water level deviation represents the fitting degree of the scheduling rules and the reservoir optimization operation process. The smaller the decision water level deviation, the stronger the fitting ability of the model, and vice versa. The power generation statistics of the scheduling rules are shown in Table 3, and the decision water level deviation of the scheduling rules is shown in Table 4. In order to present the results more clearly, the best results in the table except for the optimization scheme are highlighted in bold.

[0045] Table 3 Statistics of simulation scheduling power generation

[0046]

[0047] Table 4 Statistics of simulation scheduling decision water level deviation

[0048]

[0049] As shown in Table 3, overall, LSTM-RS has the best annual average power generation performance, which is only 0.08 million kWh less than the optimization scheme. This result preliminarily confirms the superiority of the LSTM-RS proposed in this study. Specifically, compared with ANN-RS, LGB-RS and LR-RS, the annual average power generation of LSTM-RS increases by 0.19%, 0.95% and 0.95%, respectively. Similar to this result, compared with ANN, LGB and LR, the average annual power generation of LSTM increases by 0.27%, 0.88% and 0.84%, respectively. The above results show that the rules extracted based on LSTM have better performance in power generation indicators, indicating that the LSTM model has better information extraction capability and can effectively mine the operation rules in the optimization scheme of the reservoir. In addition, compared with LSTM, ANN, LGB and LR, the power generation of LSTM-RS, ANN-RS, LGB-RS and LR-RS models increases by 0.45%, 0.54%, 0.41% and 0.33%, respectively. This result shows that the scheduling rules obtained by the rolling simulation strategy have higher power generation benefits, verifying the effectiveness of the rolling simulation strategy in extracting the scheduling rules of the reservoir.

[0050] As shown in Table 4, overall, compared with other models, the decision water level deviation of LSTM-RS is the smallest, which is only 0.66 m. This result shows that the water level simulation process of LSTM-RS is closest to the ideal solution. Specifically, compared with ANN-RS, LGB-RS and LR-RS, the decision water level deviation of LSTM-RS decreases by 7.04%, 70.67% and 69.59%, respectively. Similarly, compared with ANN, LGB and LR, the decision water level deviation of LSTM increases by 18.93%, 43.77% and 39.05%, respectively. The above results show that the rules extracted based on the LSTM model have smaller decision water level deviation, which proves the excellent fitting capability of the LSTM-RS model, which can accurately capture the water level fluctuation of the optimization scheme. In addition, compared with LSTM, ANN, LGB and LR, the decision water level deviation of LSTM-RS, ANN-RS, LGB-RS and LR-RS decreases by 60.48%, 65.47%, 24.13% and 20.73%, respectively. This result shows that the rolling simulation strategy can effectively reduce the influence of cumulative decision errors in the scheduling rules, and again verifies the effectiveness of the proposed strategy in extracting the scheduling rules of the reservoir.

[0051] For more intuitive comparison, Figure 3 The effect of the rolling simulation strategy is shown. As Figure 3As shown, in general, the water level of the model with the rolling simulation strategy is almost consistent with the ideal water level process. In the model with the rolling simulation strategy, the water level simulation effect of LSTM-RS is the best, and the water level process is closest to the ideal water level. This result is consistent with the result in Table 4, which again verifies the effectiveness of LSTM-RS in extracting the operation rules of the reservoir. In addition, the water level of the model without the rolling simulation strategy deviates significantly from the ideal water level process, especially from April to June. This is because April to June is the period when the water level of the reservoir decreases, and the water level fluctuates greatly during this period. These rules are sensitive to the change of the water level during this period, which amplifies the cumulative effect of decision-making errors, resulting in a large deviation of the decision-making water level, which indirectly verifies the superiority of the rolling simulation strategy.

[0052] In order to compare the advantages and disadvantages of the rules extracted by different models, four models, LSTM-RS, ANN-RS, LGB-RS and LR-RS, are selected for further comparison. Figure 4 The water level simulation processes of each model in different years are shown. From Figure 4 It can be seen that the water level process of LSTM-RS in each year is closest to the optimization scheme. This result again proves the superiority of the LSTM-RS model in extracting the operation rules of the reservoir. The ANN-RS model also has a relatively good water level decision-making process, but its water level process is always slightly lower than that of LSTM-RS, especially in 2019, 2021 and 2022, which affects the power generation efficiency of the reservoir. Unlike the above models, the water level processes of LGB-RS and LR-RS deviate significantly from the optimization scheme. The decision-making water level of LGB-RS and LR-RS is significantly lower in July, August and September, and the overall fitting accuracy is poor. This also indirectly explains why LGB-RS and LR-RS perform poorly in Table 3 and Table 4.

[0053] In summary, the operation rules of the reservoir extracted by the LSTM-RS model have the best overall performance. On the one hand, this result shows that LSTM has strong fitting ability and can accurately mine the potential rules in the optimized operation scheme. On the other hand, it also verifies the practicability of the rolling simulation strategy, which can effectively reduce the cumulative decision-making errors in the extraction of operation rules.

[0054] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the purpose and scope of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A method for extracting reservoir operation rules based on a rolling simulation strategy, characterized in that, The method comprises the following steps: Step 1, given the initial water level Z0 of the reservoir scheduling period, at this time the scheduling period t = 0; Step 2, with initial water level Z t and the inflow I t characteristic factor, with final water level Z t+1 target value, training model M t where M t represents the scheduling rule of the reservoir in the tth scheduling period; Step 3, get initial water level Z t and input flow I t input model M t , get decision water level K t+1 ; Step 4, judging the decision water level K t+1 whether the reservoir scheduling constraint is violated, and if so, correcting it; Step 5, decide water level K t+1 Set initial water level Z for next scheduling period t+1 and let t = t + 1; Step 6: judging whether the scheduling period t meets the end condition, if yes, outputting the final scheduling rule, otherwise, cyclically executing steps 2-6.

2. The method of claim 1, wherein the method is characterized by: In Step 3, the decision water level K t+1 is the model M t The final water level predicted by the model, and the final water level Z t+1 There is a certain deviation value.

3. The method of claim 1, wherein the method is characterized by: In step 4, the specific method of correction is that if the decision water level exceeds the upper limit value, the decision water level is forced to be equal to the upper limit value, and if the decision water level exceeds the lower limit value, the decision water level is forced to be equal to the lower limit value.

4. The method of claim 1, wherein the method is characterized by: In steps 2 and 3, the model comprises a long short-term memory neural network model, an artificial neural network model, a light gradient boosting machine model and a linear regression model.

Images