Multi-energy complementary coordinated scheduling method based on improved step-by-step optimization algorithm

Abstract

The present invention discloses a multi-energy complementary coordinated scheduling method based on an improved step-by-step optimization algorithm, belonging to the technical field of multi-energy complementary scheduling of water, wind, solar and storage. The method comprises: obtaining hydrological, engineering and load data to generate a preliminary feasible scheduling solution; using an improved step-by-step optimization algorithm (POA) for iterative search, wherein the improvement comprises adopting a variable discrete step length, a dynamic feasible domain and a combined heuristic candidate update strategy; combining the output sequences of each subsystem, introducing a dynamic credit value based on nonlinear updates to perform weighted redistribution of load deviations; and finally evaluating and selecting the solution with the greatest peak-shaving benefit through a scheduling objective function coupled with a dynamic penalty function. By introducing a credit mechanism, a gradient dead zone solution and the like, the present invention significantly improves the intelligence level of scheduling, computational stability and the economic benefits of the final solution.

Description

Technical Field

[0001] The present invention relates to a scheduling algorithm, in particular to a multi-energy complementary collaborative scheduling method based on an improved step-by-step optimization algorithm. Background Art

[0002] Renewable energy sources such as wind and solar power are characterized by significant intermittency, volatility, and randomness. Their large-scale grid integration presents unprecedented challenges to the safe and stable operation of power systems. The coordinated and complementary dispatch of multiple energy sources, including hydropower, wind power, solar power, and storage, is a key technical means to mitigate fluctuations in renewable energy output and enhance system flexibility and reliability. It holds significant practical value for ensuring the safe and stable operation of power grids, improving energy efficiency, and promoting the transition to a cleaner energy structure. Hydropower, as the most regulative clean energy source, plays a core role as a stabilizer and regulator in this multi-energy complementary system. Therefore, developing optimized dispatching methods that efficiently and precisely coordinate the combined operation of hydropower, wind power, solar power, and storage, and fully tapping the peak-shaving potential of hydropower to support renewable energy consumption, is a critical and pressing issue in the field of power system optimization and dispatch, directly impacting the development of new power systems and the level of energy security.

[0003] Currently, algorithms for solving the coordinated scheduling of multi-energy complementary systems, including hydropower, can be categorized into two main technical systems: dynamic programming (DP) and its derivatives, and heuristic algorithms (HA). Based on rigorous mathematical optimization theory, DP algorithms decompose the full scheduling cycle problem into a series of single-stage subproblems for solution. They demonstrate theoretical advantages in addressing scheduling problems characterized by Markov decision processes. Among them, the stepwise optimization algorithm (POA), a widely used DP approximation algorithm, effectively alleviates the curse of dimensionality encountered by traditional DP in high-dimensional problems by iteratively searching for optimal solutions within a given initial feasible solution corridor. On the other hand, heuristic algorithms, such as genetic algorithms (GAs) and particle swarm optimization (PSOs), are rooted in biomimetic or physical mechanisms. By simulating evolutionary or swarm collaboration processes in nature, they conduct a randomized, global search across the entire solution space. These algorithms have low requirements for the mathematical properties of the problem model and exhibit good generalizability and global optimization capabilities when handling complex scheduling models with nonlinear, non-convex, and multi-constrained characteristics. They can provide high-quality, feasible scheduling solutions for large-scale systems within a reasonable timeframe.

[0004] While these algorithms have been successfully applied in specific scenarios, existing technologies still face challenges in accuracy, efficiency, and intelligence when faced with the new challenges of tighter multi-energy coupling and increased uncertainty in new power systems. For example, existing technologies face challenges such as insufficient numerical stability in critical operating ranges and rigid load allocation mechanisms in multi-agent collaboration. Summary of the Invention

[0005] Purpose of the Invention: This invention aims to address technical issues with existing scheduling methods when dealing with high-dimensional, nonlinear problems, such as numerical computation instability, rigid load allocation mechanisms, and low solution efficiency and quality. It provides a multi-energy complementary collaborative scheduling method based on an improved stepwise optimization algorithm.

[0006] The technical solution is a multi-energy complementary coordinated scheduling method based on an improved step-by-step optimization algorithm, including:

[0007] Obtain hydrological data, engineering data and total grid load;

[0008] Generate a preliminary feasible dispatch solution used as an initial solution based on hydrological data and total grid load;

[0009] An improved stepwise optimization algorithm is used to iteratively search for the preliminary feasible scheduling solution to generate the output sequence of each subsystem.

[0010] Combine the output sequences of each subsystem and redistribute the load according to the deviation from the total load of the power grid to generate an optimized output sequence;

[0011] Evaluate the optimized output sequence and select the scheduling plan with the greatest peak-shaving benefit.

[0012] Beneficial effects: By introducing a credit mechanism, a gradient dead zone solution, dynamic allocation based on the rate of change of water levels, and a dynamic penalty function coupled with an objective function, this invention significantly improves the intelligent level of scheduling, computational stability, and the economic benefits of the final solution. These technical effects are described in detail below in conjunction with specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of the present invention.

[0014] Figure 2 It is a flow chart of the present invention for generating an optimized output sequence.

[0015] Figure 3 The present invention generates a flowchart of a preliminary feasible scheduling solution used as an initial solution through iterative search.

[0016] Figure 4 It is a flow chart of generating an adaptive elastic coefficient according to the present invention.

[0017] Figure 5 It is a flow chart of the flow distribution process of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following Figures 1 to 5, and specific embodiments are provided to further describe the present invention in detail. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0019] In order to solve the above problems existing in the background technology, in-depth analysis and research were carried out:

[0020] First, the problem of numerical calculation stability. In the process of solving the power generation flow by reverse engineering the power generation flow, the existing gradient-based optimization method will encounter a gradient dead zone under specific hydraulic conditions. Specifically, when the reservoir is at a high water level or has a strong discharge capacity, the sensitivity of the power generation head to the change in power generation flow (i.e., the head gradient dH / dQ) will become extremely low. At this time, the algorithm with gradient as the search direction will stall due to weak or disappearing gradient information, causing the iterative process to stagnate or oscillate violently near the optimal solution, and unable to converge to the true optimal operating point. This problem is not simply a lack of calculation accuracy, but stems from the failure of the algorithm mechanism within a specific physical range, which makes it impossible to provide reliable and repeatable optimization results for scheduling decisions, especially in critical scheduling scenarios such as flood control season or high power generation efficiency areas. This defect may lead to huge economic losses or safety risks.

[0021] Secondly, the problem of rigid load allocation mechanisms is addressed. In the coordinated dispatch of multiple hydropower stations, when the combined total output deviates from grid demand, existing load redistribution strategies typically employ static ratios based on fixed physical parameters such as installed capacity. This one-size-fits-all approach completely ignores the actual current and recent operating conditions and historical performance of each hydropower station. For example, when undertaking emergency regulation tasks, existing mechanisms often treat a hydropower station that maintains an efficient and flexible state (e.g., near the desired load factor) by accurately responding to dispatch instructions indiscriminately, while another hydropower station deviates from its economic operating range due to unstable output. This mechanism lacks dynamic adaptability and fairness, failing to provide positive incentives for high-performing hydropower stations. In the long run, it also inhibits the self-optimization initiative of various dispatching entities. This, in turn, makes it difficult to achieve optimal synergy and resource utilization efficiency across the entire system, ultimately failing to achieve truly intelligent dispatch.

[0022] To this end, the following implementation plan is provided. It should be noted that the hydropower, wind power, photovoltaic power generation, and pumped storage multi-energy complementary system mentioned in the embodiments of the present invention refers to a power system that includes multiple energy forms such as hydropower, wind power, photovoltaic power, and pumped storage, and coordinates and dispatches them in a unified and coordinated manner. This is intended to address the peak-shaving difficulties and power balance challenges brought about by the volatility and intermittency of renewable energy in short-term scheduling of such systems.

[0023] Example 1 describes the overall process of a multi-energy complementary coordinated scheduling method based on an improved step-by-step optimization algorithm. This method reduces and decouples complex scheduling problems by constructing multiple closed-loop feedback mechanisms, efficiently generating a scheduling solution that balances peak-shaving efficiency and operational safety. Specifically, it includes the following steps:

[0024] The first step is to obtain input data and perform system dimensionality reduction processing.

[0025] In the specific implementation process, hydrological data, engineering data and total power grid load are first obtained.

[0026] Hydrological data includes inflows to each reservoir within the dispatch cycle, interval inflows, and ecological baseflow requirements. This data is the foundation for water balance calculations and power generation capacity forecasts. Engineering data includes the installed capacity, guaranteed output, water level-reservoir capacity curve, and water level-discharge capacity curve of each hydropower station; the installed capacity and efficiency coefficient of pumped-storage power stations; the installed capacity and efficiency coefficient of new energy power stations (wind power, photovoltaic power); and the corresponding total grid load at each time step. The total grid load refers to the time series of all electricity demands that the power system must meet within the dispatch cycle.

[0027] This step adopts the concept of large-scale system decomposition and coordination. First, the hydropower, wind, solar, and storage multi-energy complementary system is spatially divided into multiple independent subsystems. For example, each hydropower station and each new energy station can be considered an independent subsystem or entity. Subsequently, the grid load of each subsystem at the same time is aggregated in the time dimension to form the total load curve of the virtual aggregated grid. Through this hierarchical and block-based modeling approach, the high-dimensional complex system scheduling problem is reduced to the collaborative optimization problem of multiple subsystems, laying the foundation for subsequent efficient solutions.

[0028] Step 2: Generate a preliminary feasible dispatch solution used as an initial solution. Specifically, after acquiring data and reducing dimensionality, this method generates a preliminary feasible dispatch solution used as an initial solution based on hydrological data and total power grid load.

[0029] The quality of traditional optimization algorithms depends heavily on the selection of the initial solution. A high-quality initial solution can significantly accelerate the algorithm's convergence and prevent it from getting stuck in a local optimum. Using a closed-loop iterative approach based on the principle of determining water flow by electricity, an initial solution is generated that closely aligns with the target load and satisfies basic physical constraints, providing an excellent starting point for subsequent refined searches.

[0030] The core of this step is to invert the operating state of the hydropower station required to meet the load based on the target load through a closed-loop iterative process of head-flow-water level, thereby achieving decoupling of the actual output and the target load.

[0031] Specifically, the system first determines whether the target load assigned to a hydropower station is greater than the guaranteed output. If normal power generation is possible, starting from the initial power generation flow (such as ecological base flow), the head loss is fitted, the head and actual output are iteratively calculated, and compared with the target load. During the iteration process, the system will dynamically adjust the division of power generation flow and abandoned water flow, and make judgments based on the dual convergence criteria of water level and flow until the calculated actual output converges to the target load. At the same time, this process always checks the flood control limit water level, maximum downstream flow, ecological base flow, etc. as hard constraints to ensure that the generated preliminary solution is physically feasible. Subsequent Examples 2 and 3 will elaborate on the two preferred technical solutions for generating preliminary feasible solutions.

[0032] Step 3: Use the improved stepwise optimization algorithm for iterative search. Specifically, based on the high-quality preliminary feasible scheduling solution, this method uses the improved stepwise optimization algorithm to iteratively search the preliminary feasible scheduling solution to generate the output sequence of each subsystem.

[0033] The stepwise optimization algorithm is a dynamic programming algorithm commonly used to solve multi-stage decision-making problems. Its basic idea is to optimize only the decision variables for the current period, fixing the decision variables for all other periods. The algorithm then iterates through all periods until the results converge.

[0034] The present invention has made many improvements to the traditional POA algorithm. First, a compact dynamic feasible domain is defined based on the preliminary feasible solution and physical constraints, and all subsequent searches are performed within this domain, avoiding invalid calculations. Secondly, when discretizing the water level at each moment to find the optimal decision, a variable discrete step size strategy is adopted. In the early stage of iteration, a large step size is used for global exploration, and in the later stage, a small step size is used for refined search, taking into account both globality and convergence accuracy. Finally, the traditional POA distributed search is combined with the heuristic candidate update processing objective function to update the decision variables. The detailed implementation of this step will be further explained in Example 5.

[0035] Step 4: Load redistribution and output sequence optimization.

[0036] After obtaining the output sequence of each subsystem, this method combines the output sequences of each subsystem and performs load redistribution based on the deviation between the output sequence and the total load of the power grid to generate an optimized output sequence.

[0037] First, the output sequences of each subsystem (hydropower, wind power, photovoltaic power, etc.) generated in the previous step are superimposed at each moment to obtain the system's total output curve. Then, the deviation between this total output curve and the actual total grid load curve is calculated. This deviation represents the load to be redistributed. This invention proposes a tiered load gap filling strategy. Based on whether the total output is sufficient or insufficient, this deviation is redistributed using different capacity proportional coefficients, and the output of each hydropower station is updated. This step effectively smooths fluctuations in the system's total output, allowing it to more closely track load demand. The following Example 4 will describe a preferred technical solution for load redistribution (based on a credit mechanism).

[0038] Step 5: Evaluate and select the final solution.

[0039] This method evaluates the optimized output sequence and selects the scheduling scheme with the greatest peak-shaving benefit.

[0040] The core of this step is to transform the multi-objective optimization problem into a single-objective optimization problem coupled with the power balance constraint. Specifically, an objective function is established with the maximization of total power generation revenue (or minimization of water abandonment) as the main objective, and a dynamic penalty function term is embedded. The penalty function is positively correlated with the degree of deviation of the total output of the system from the grid load. When the scheduling scheme seriously violates the power balance constraint, its objective function value will be significantly penalized, and thus it will be naturally eliminated in the optimization process. Furthermore, the adjustment information of the constraint violation will be fed back to the POA iterative loop in the third step in real time, and its core variables will be dynamically adjusted to form a closed-loop feedback, thereby fundamentally guiding the search towards the direction of satisfying the power balance. The subsequent Example 6 will elaborate on this evaluation mechanism in detail.

[0041] Through the organic combination of the above five steps, the present invention can provide an efficient, reliable and most cost-effective coordinated scheduling solution for complex water, wind, solar and storage multi-energy complementary systems.

[0042] Embodiment 2 describes a preliminary solution generation method designed to solve the gradient dead zone problem during the process of generating a preliminary feasible scheduling solution.

[0043] During the iterative search process, the method calculates the hydraulic head sensitivity of the current hydraulic head to the current power generation flow.

[0044] When the hydraulic head sensitivity falls below a preset sensitivity threshold, a virtual potential function is constructed, linking the current power flow rate with the target flow rate. The gradient of the virtual potential function is then calculated and combined with the hydraulic head sensitivity to form a revised search gradient. Finally, the revised search gradient is used to update the current power flow rate until convergence leads to a preliminary feasible dispatch solution.

[0045] Head sensitivity refers to the rate of change of reservoir head (a key component of power generation head) relative to the generated flow rate, or dH / dQ. In certain reservoir operating ranges, changes in generated flow rate have little impact on head, causing this sensitivity to approach zero. A gradient dead zone refers to a phenomenon in which, in a gradient-based optimization algorithm, the gradient (or sensitivity) of the objective function with respect to the decision variable is too small, preventing the algorithm from determining a valid search direction and causing iterations to stagnate.

[0046] The virtual potential function is an auxiliary function introduced to solve the gradient dead zone problem. This function constructs a specific potential field in the solution space, and the potential gradient it generates can guide the search direction out of the dead zone.

[0047] In reservoir operation calculations, if the search relies solely on head sensitivity as the gradient, the algorithm will fail once it enters the gradient dead zone. By introducing a virtual potential energy function pointing to the target flow and combining its gradient with the original gradient, an external traction force is applied when the search stagnates, pulling the search process out of the flat zone and ensuring the stability and efficiency of the iterative process.

[0048] In each iteration, the sensitivity dH / dQ of the current water head H to the power generation flow Q is calculated. The sensitivity threshold εH is set (for example, 10 -4 ). When |dH / dQ|<εH, the virtual potential energy mechanism is triggered.

[0049] Construct a quadratic virtual potential energy function U virtual (Q)=0.5k(QQ target ) 2 Among them, Q target It is the theoretical power generation flow calculated based on the target output that the current hydropower station needs to bear; k is the adaptive elastic coefficient.

[0050] Search gradient correction and update: Calculate the gradient of the virtual potential energy function grad U virtual =k×(QQ target ). This gradient is weightedly combined with the original head sensitivity to form the modified search gradient G modified =w1×(dH / dQ)+w2×grad U virtual 。Where w1 and w2 are weight coefficients. Use this G modified Update the current power generation flow Q as the direction.

[0051] The convergence of iteration adopts double convergence criterion, that is, when the water level Z calculated in this round of iteration is t (i.e., water level at time t) and power generation flow Q t , and the corresponding value Z calculated in the previous iteration t-1 and Q t-1When the changes between them are all less than their respective preset convergence thresholds, the iteration is determined to have converged.

[0052] Furthermore, to improve the effect, the virtual potential energy function includes an adaptive elastic coefficient. The generation method thereof includes:

[0053] Based on the hydraulic head sensitivity, the initial value of the adaptive elastic coefficient is determined, where the initial value is exponentially inversely proportional to the hydraulic head sensitivity. For example, it can be expressed as k0 = A × exp(-B × |dH / dQ|), where A and B are positive constants. The smaller the gradient, the stronger the guiding force applied.

[0054] During the iterative search process, the updated amount of power generation flow is tracked and recorded to form a flow update amount sequence.

[0055] Based on the latest power generation flow update and the changing trend of the flow update sequence, the adaptive elastic coefficient is dynamically attenuated or enhanced. For example, when the flow update direction is stable, the k value can be appropriately reduced to perform a more refined search; when the flow update is oscillating, the k value can be increased to strengthen the guidance.

[0056] Example 3: Describes a smooth traffic allocation method based on a continuous function in the process of generating a preliminary feasible scheduling solution.

[0057] In the specific implementation process, the process of generating a preliminary feasible scheduling solution further includes the traffic allocation process, specifically:

[0058] Based on the historical water level series and the current water level, calculate the water level change rate at the current moment.

[0059] Based on the rate of change of the water level, the successive allocation ratios for dividing the total available flow are calculated.

[0060] The total available flow is smoothly divided into power generation flow and abandoned water flow using a continuous allocation ratio.

[0061] In traditional scheduling, decisions between power generation and water abandonment are often made in a discrete, either-or-not model. This can easily cause drastic changes in scheduling results at critical points (such as when the flood control limit water level is about to be reached), which is detrimental to system stability. This embodiment introduces a continuously varying allocation ratio within the interval [0, 1] to achieve a smooth transition of the total available flow between power generation and water abandonment, effectively avoiding system fluctuations caused by discrete decisions and improving scheduling stability and robustness.

[0062] For example, the process is as follows:

[0063] First, the water level margin is calculated based on the current water level and the flood control limit water level. Specifically, the water level margin γ = Z flood -Z current , where Zflood is the flood control limit water level, Z current is the current reservoir water level. This margin reflects the reservoir's current flood control safety. Furthermore, the current water level change rate, dZ / dt, is calculated based on historical water level data. This rate reflects the fluctuation trend of the reservoir water level.

[0064] The water level change rate and water level margin are used as inputs to generate a continuous allocation ratio through the hyperbolic tangent function.

[0065] The continuous allocation ratio α is a variable that takes continuous values ​​in the interval [0, 1] and is used to determine the total available flow Q. total The proportion of electricity used for power generation.

[0066] The hyperbolic tangent function (tanh) is an S-shaped activation function with a range of (-1, 1) and a smooth transition characteristic, making it very suitable for generating continuous proportions.

[0067] The calculation formula of the ratio α can be designed as: α=0.5+0.5×tanh(c1×(dZ / dt)+c2×γ); or α=0.5+0.5×tanh(c×(dZ / dt)×γ); where c, c1, and c2 are adjustment parameters set according to actual engineering experience.

[0068] Traffic smoothing division: The total available traffic Qtotal is divided using the calculated continuous allocation ratio α:

[0069] Power generation flow Q gen =α×Q total ; Discarded water flow Q spill =(1-α)×Q total ;

[0070] Due to the characteristics of the hyperbolic tangent function, when the water level rises rapidly (dZ / dt is large) or approaches the flood control limit (γ is small), α will smoothly approach 0, thereby discharging most of the water and ensuring flood control safety. Conversely, when the water level is stable or falling, α will approach 1, giving priority to meeting power generation needs.

[0071] Embodiment 4 describes a method based on a nonlinear credit mechanism in the load redistribution link.

[0072] In the specific implementation process, the output sequences of each subsystem are combined and load redistribution is performed based on the deviation between the output sequence and the total load of the power grid to generate the optimized output sequence. The process specifically includes:

[0073] Obtain the current credit value of each reservoir and, combined with the installed capacity in the project data, calculate the credit-weighted allocation weight for each reservoir. Determine the total load to be redistributed based on the deviation between each subsystem's output sequence and the total grid load. Use the credit-weighted allocation weights to allocate the total load to be redistributed, and generate an optimized output sequence.

[0074] When redistributing system output deviations, allocating them solely based on fixed proportions, such as installed capacity, fails to incentivize optimal performance among power plants. This embodiment introduces a dynamic credit mechanism, using credit values ​​to quantitatively assess each reservoir's historical performance (e.g., load saturation). When allocating tasks, power plants with higher credit ratings and better historical performance are assigned more regulation tasks, achieving fairer and more efficient load redistribution and establishing a self-regulating and long-term stable positive incentive system.

[0075] For example, the process is as follows: First, obtain the credit value C accumulated by each reservoir i at the end of the previous scheduling cycle i Combined with its installed capacity N max，i , calculate its credit weighted allocation weight W i Preferably, the weight can be calculated as follows: i =C i ×N max，i / ∑ j (C j ×N max，j ); where the denominator is the sum of the credit and capacity products of all reservoirs participating in the regulation, j is the total number of reservoirs, C j is the credit value accumulated by each reservoir j at the end of the previous scheduling cycle, N max,j is the installed capacity of the corresponding regulating reservoirs.

[0076] Calculate the deviation between the total output of all subsystems and the total load of the power grid to obtain the total load to be redistributed ΔL (negative if the total output is insufficient, positive if it is sufficient). i Assigned to each reservoir i, the adjustment load it needs to bear is ΔL i =ΔL×W i The adjustment amount is added to the original output of reservoir i to obtain the optimized new output, completing the load redistribution process.

[0077] Assume that reservoir A and reservoir B participate in load redistribution.

[0078] Initial credit value C A =100, C B =120 (Reservoir B has better historical performance).

[0079] Installed capacity N max，A=500MW, N max，B =300MW.

[0080] The total load to be redistributed (insufficient output) ΔL=-80MW.

[0081] The calculation process is as follows:

[0082] Calculate the product of credit and capacity: A: 100 × 500 = 50,000; B: 120 × 300 = 36,000;

[0083] Calculate the credit-weighted allocation weight: WA = 50,000 / (50,000+36,000) ≈ 0.581; WB = 36,000 / (50,000+36,000) ≈ 0.419;

[0084] Distributed adjustment load: ΔLA=-80MW×0.581≈-46.48MW; ΔLB=-80MW×0.419≈-33.52MW.

[0085] Reservoir A needs to increase its installed capacity by 46.48 MW, while Reservoir B needs to increase its installed capacity by 33.52 MW. Therefore, although Reservoir B has a smaller installed capacity, it receives more weight in addition to the allocation based on pure capacity due to its higher credit value.

[0086] In order to make the credit mechanism dynamic and long-term stable, the present invention further includes a nonlinear credit update process:

[0087] Based on the optimized output sequence and the rated capacity in the engineering data, the current load rate of each reservoir is calculated.

[0088] The credit increment is calculated according to the deviation between the current load rate and a preset expected load rate, wherein the credit increment is proportional to the square of the deviation.

[0089] The credit increment is added to the current credit value of each reservoir to generate an updated credit value for use in the next scheduling cycle.

[0090] Specific implementation: current load rate η i =Optimized output i / N max，i Credit increment ΔC i =β×(η i -η exp ) 2 ×sign(η i -η exp ). Where β is the adjustment coefficient, η exp is the expected load rate (e.g. 0.8), sign() is the sign function. The quadratic term amplifies the reward and penalty effects. The updated credit value is C i，new =C i +ΔCi .

[0091] In order to prevent the credit value from becoming extreme due to short-term fluctuations, the present invention also maintains a credit value history sequence including the historical credit value of each reservoir.

[0092] Add the updated credit value to the historical credit value series and calculate the volatility indicators of the series (such as standard deviation).

[0093] When the volatility indicator exceeds the preset volatility threshold, a mean reversion force toward the benchmark credit value is applied to the updated credit value to generate a balanced adjusted credit value, which is used to replace the updated credit value for use in the next scheduling cycle.

[0094] Specifically, when volatility exceeds a threshold, a mean reversion force is applied: C i，final =C i，new -λ×(C i，new -C base ). Where λ is the regression coefficient (such as 0.1), C base It is the benchmark credit value (e.g. the initial value is 100). This ensures the long-term stability of the credit system.

[0095] Example 5: Describe the implementation process of the improved POA algorithm.

[0096] In the specific implementation process, the improved stepwise optimization algorithm is used to iteratively search for the preliminary feasible scheduling solution, including: based on the preliminary feasible scheduling solution and combined with the physical constraints in the engineering data, the dynamic feasible domain is delineated.

[0097] The process of generating an initial water level in the dynamic feasible region, which is used as the starting point of the iterative search.

[0098] In the dynamic feasible region, the water level at a certain moment in the initial water level process is discretized to generate candidate water levels.

[0099] The discretization process adopts a variable discrete step size strategy, where the step size is determined according to the current number of iterations and the logarithmic decay law.

[0100] Traditional POA algorithms have two major issues: first, the randomness of the initial solution can lead to inefficient search; second, using a fixed discrete step size makes it difficult to strike a balance between global exploration and local accuracy; and third, distributed search combined with heuristic candidate updates to process the objective function and update decision variables. This embodiment, through three core improvements: defining a dynamic feasible region, logarithmically decaying step sizes, and heuristic candidate update decision variables, ensures that the search remains in a favorable region. It automatically adjusts the search granularity at different stages of the iteration and guides updates during the search process, significantly improving search efficiency and the accuracy of the final solution while ensuring global optimality.

[0101] Specifically, it includes the steps of delineating the dynamic feasible domain, discretizing with logarithmic decay variable step size, and updating the decision variables of the objective function by combining distributed search with heuristic candidate update.

[0102] Rather than starting with an arbitrary water level process, this approach first uses the preliminary feasible solution generated in Example 1, combined with physical constraints such as the reservoir's maximum and minimum operating water levels, to define a more compact, time-varying dynamic feasible domain. All subsequent water level searches and discretization operations are strictly confined to this feasible domain, ensuring the feasibility of intermediate solutions from the outset and avoiding significant inefficient computations.

[0103] When optimizing the water level at a certain time t, a set of discrete candidate water levels needs to be generated. This method adopts a variable step size strategy. Let the current iteration be the jth iteration and the total number of iterations be J max , then the discrete step length ΔZ of this iteration j Can be set to: ΔZ j =ΔZ end +(ΔZ start -ΔZ end )×(1-log 10 (1+9×j / J max )). A more concise statement is that the step size decays logarithmically with the number of iterations j. start is the larger initial step size, ΔZ end is the smaller final step size.

[0104] At the beginning of the iteration (j is small), the step size ΔZ j Close to ΔZ start , allowing the algorithm to conduct large-scale, coarse-grained exploration within the feasible domain to find the global optimal area.

[0105] As the iteration progresses (j increases), the step size ΔZ j It decreases smoothly according to the logarithmic curve and approaches ΔZ end The algorithm guides the update through the objective function, starting from the current optimal solution, taking the dynamic feasible region as the boundary, and adjusting the outflow heuristically according to the "expected load-actual output" relationship; updates are accepted only when the objective function is improved, so that the algorithm can perform small-scale, refined searches in more promising areas to improve the accuracy of the final solution.

[0106] Example 6: Describe the solution evaluation and selection process based on dynamic penalty function and real-time feedback.

[0107] In the specific implementation process, the process is as follows:

[0108] The steps of evaluating the optimized output sequence and selecting the dispatching scheme with the greatest peak-shaving benefit include: establishing a multi-energy complementary power balance constraint, which is used to measure the consistency between the optimized output sequence and the total load of the power grid.

[0109] According to the power balance constraint: L s (t)=∑[∑N i (t)+∑N cx，j (t)+∑N x，k (t)], where L s (t) is the actual total power generation output in period t, which is compared with the grid load demand L grid (t) violation degree, construct dynamic penalty function. ∑N i (t) is the total output of the hydropower station in time period t, ∑N cx,j (t) is the total power output (generated) by all pumped storage power stations in time period t, ∑N x,k (t) is the total output of new energy (wind, solar) power stations in time period t.

[0110] The dynamic penalty function is coupled to the scheduling objective function, and the evaluation and selection are completed by solving the scheduling objective function. Preferably,

[0111] When the optimized output sequence violates the power balance constraint, the adjustment information used to satisfy the constraint is synchronously fed back to the iterative loop of the improved step-by-step optimization algorithm, and the scheduling core variables within the loop are updated.

[0112] Multi-energy complementary scheduling is essentially a multi-objective optimization problem that requires strict compliance with power balance. Traditional methods usually penalize solutions that violate constraints only after all searches are completed, which is inefficient. This embodiment transforms complex power balance constraints into part of a single-objective optimization problem through a dynamic penalty function, and performs real-time processing during the algorithm's iteration process. More importantly, it establishes a closed-loop mechanism of evaluation and feedback, allowing the POA algorithm to perceive downstream constraints during the search and actively adjust its direction, thereby avoiding wasting computing resources in the solution space with serious violations, greatly improving the algorithm's convergence speed and the quality of the final solution.

[0113] Specifically, the process is as follows:

[0114] First, establish the power balance constraint, that is, the total output of the system at time t (the sum of hydropower, wind power, solar power, storage, etc.) and the total load of the grid L grid The deviation δ(t) of (t) should be as small as possible. According to the degree of violation of the deviation, a dynamic penalty function is constructed, such as P(t)=w(t)×|δ(t)| 2 , where w(t) is the dynamic weight coefficient that is adaptively adjusted during the iterative process.

[0115] The penalty function term is added to the original dispatch objective function. For example, if the original objective is to maximize the total power generation F gen =∑[∑N i (t)+∑N cx，j (t)+∑N x，k (t)], where ∑N i (t) is the sum of the power generation output of all reservoirs in time period t; ∑N cx，j (t) is the sum of the discharge (generation) output of all pumped storage power stations in time period t; ∑N x，k (t) is the sum of the power output of all new energy power stations in period t; then the new objective function after coupling F obj Change to F obj =F gen -∑P(t). In solving F obj In the process of maximization, any solution that causes δ(t) to deviate from 0 will cause F to obj It decreases and is automatically suppressed by the optimization process.

[0116] It's important to note that during the POA iteration process, if an intermediate candidate solution causes an output deviation δ(t) at time t, the system doesn't wait until the end to penalize it. Instead, it immediately calculates the output adjustment required to eliminate this deviation. This adjustment information (for example, the required output increase or decrease for a particular reservoir) is immediately fed back into the POA algorithm's iteration loop and used to directly modify the core scheduling variables currently being optimized (such as candidate water levels or power flow), thereby guiding the next step of the search. This real-time closed-loop feedback forms the fundamental connection between power demand, water level regulation, and flow control, ensuring that the algorithm consistently conducts efficient exploration around the core constraint of satisfying power balance.

[0117] Example 7: Demonstrates the specific calculation process of dynamic load redistribution based on a nonlinear credit mechanism.

[0118] Suppose there are two hydropower stations in a power system: Reservoir A and Reservoir B.

[0119] Among them, reservoir A: installed capacity N max_A =300MW. Current credit value C A = 110. Reservoir B: installed capacity N max_B =200MW. Current credit value C B =95.

[0120] System parameters include expected load rate eta exp =0.8; credit incremental adjustment coefficient beta=0.5; mean regression coefficient lambda=0.1; benchmark credit value C Base =100.

[0121] In a certain dispatching period t: the initial optimization result of the POA algorithm is: Reservoir A output N A_poa =230MW, Reservoir B output N B_poa =160MW. Total output N total_poa =230+160=390MW. The actual grid load during this period, L(t), is 400MW. At this point, there is a load shortfall of delta(t) = 390-400 = -10MW. This -10MW is the total load delta(t) to be redistributed.

[0122] The calculation process is as follows:

[0123] Step 1: Calculate the credit-weighted allocation weights

[0124] Weighted cardinality of reservoir A = C A *N maxA =110*300=33000. The weighted base number of reservoir B = C B *N maxB =95*200=19000. Total weighted base =33000+19000=52000. Allocation weight W of reservoir A A =33000 / 52000≈0.635. The allocation weight W of reservoir B B =19000 / 52000≈0.365.

[0125] Step 2: Perform load redistribution

[0126] The adjustment amount that reservoir A needs to bear is delta A =delta(t)*W A =10MW*0.635=6.35MW. (Because it is a gap, the output is increased); the adjustment amount that reservoir B needs to bear is delta B =delta(t)*W B =10MW*0.365=3.65MW.

[0127] Optimized output sequence: Reservoir A output N after optimization A_final =230+6.35=236.35MW. Output of reservoir B after optimization N B _final =160+3.65=163.65MW.

[0128] N A_finall +N B _final =236.35+163.65=400MW, which has been verified to be consistent with the grid load.

[0129] Step 3: Calculate credit increment

[0130] Calculate the new load rate: Reservoir A load rate etaA =236.35 / 300≈0.788. Reservoir B load rate eta B =163.65 / 200≈0.818.

[0131] Calculating credit increment: Delta C_A =beta*(eta A -eta exp )2*sign(eta A -eta exp )=0.5*(0.788-0.8) 2 *sign(0.788-0.8)=0.5*(-0.012) 2 *(-1)=-0.000072. (Credit slightly decreased due to slightly lower than expected load factor); Delta C_B =beta*(eta B -eta exp ) 2 *sign(eta B -eta exp )=0.5*(0.818-0.8) 2 *sign(0.818-0.8)=0.5*(0.018) 2 *(1)=0.000162. (Credit increased slightly due to slightly higher than expected load factor).

[0132] Step 4: Generate updated credit value for next cycle

[0133] Updated credit value C A_new =C A +Delta C_A =110-0.000072=109.999928.

[0134] Updated credit value C B_new =C B + Delta C_B =95+0.000162=95.000162.

[0135] Step 5 (optional): Check whether mean reversion is necessary to reduce C A_new and C B_new Add their respective historical credit series and calculate volatility indicators (such as standard deviation). Assuming that the calculated volatility does not exceed the preset threshold, no mean reversion force will be applied in this cycle. The next scheduling cycle will directly use C A_new and C B_new As the initial credit value. If it exceeds the threshold, mean reversion calculation is required.

[0136] In summary, in this application, by establishing a new load distribution paradigm that can quantify historical performance and has dynamic incentives and long-term stability characteristics, the intelligence level, fairness and overall operational efficiency of the multi-agent collaborative dispatching system are significantly improved.

[0137] Specifically, by nonlinearly (quadratically) correlating credit values ​​with the deviation between the load rate and the expected load rate, the system accurately quantifies the operating performance of each power plant, clearly distinguishing between rewards and penalties. Even minor performance differences are amplified, creating a strong positive incentive. Secondly, when redistributing load, this credit value is combined with installed capacity for credit weighting, ensuring that power plants with excellent historical performance and strong regulation capabilities can take on more important regulation tasks, achieving a leap from static allocation to performance-based allocation. Finally, the introduction of mean reversion to achieve long-term balanced adjustment of credit values ​​prevents the unlimited accumulation or decay of credit values ​​for some power plants, ensuring the long-term health and stability of the entire credit system.

[0138] In the scenario of joint operation of multiple cascade reservoirs, this mechanism not only achieves load tracking through secondary optimization of load distribution, but also avoids the "whipping the bull" phenomenon caused by the use of static capacity ratio allocation. For example, a reservoir that has made significant contributions to the system's peak and valley filling through refined operation will have its credit value increased accordingly. When the system encounters a load gap again, it will receive a more reasonable additional issuance task due to its high credit weight, rather than being unfairly allocated too much load. This not only protects the healthy operation of individual power plants but also, by incentivizing self-optimization among various entities, ultimately improves the water resource utilization efficiency of the entire river basin and the overall power generation efficiency of the system.

[0139] By constructing an adaptive virtual force field in the iterative algorithm, the problem of numerical calculation failure of the traditional gradient optimization method under specific hydraulic conditions is innovatively solved, ensuring the stability and convergence of the algorithm under the entire working condition range.

[0140] Specifically, this mechanism is achieved through a dual effect. First, when the algorithm detects that the head gradient dH / dQ is lower than the threshold (that is, it enters the gradient dead zone), it will actively construct a virtual potential energy function. This function can artificially form a potential well at the target flow point, and its gradient provides the algorithm with a clear and effective search direction when the physical gradient disappears. Second, the intensity of the virtual potential energy (that is, the adaptive elastic coefficient k) is dynamically adjustable. Its size is exponentially inversely proportional to the physical gradient dH / dQ, which means that the weaker the physical gradient, the stronger the virtual guiding force, achieving precise compensation of weak and strong. This adaptive mechanism ensures that the guiding force is just right, which can effectively pull out the dead zone without excessively interfering with the normal gradient search process.

[0141] In real-world scenarios like high water levels during flood season or when joint reservoir scheduling causes downstream water level support, the hydraulic head of a hydropower station is highly insensitive to flow rate changes. Traditional models that determine water flow based on power generation are prone to failure or erroneous results in these situations. This invention ensures that even under these extreme but common operating conditions, the scheduling model can still stably and quickly converge to an accurate power generation flow rate. This provides reliable computational support for flood control decisions and the development of power generation plans within the economically efficient range, avoiding the significant economic losses caused by model failure.

[0142] By introducing a continuous allocation ratio based on the hyperbolic tangent function, the discrete decision-making process of power generation and water abandonment is smoothed, which effectively suppresses the output oscillation of the scheduling model during critical state switching and improves the stability and physical rationality of the scheduling scheme.

[0143] Specifically, a continuous function replaces the traditional if-then discrete logic. Traditional water abandonment decisions are 0-1 problems: either full power generation or water abandonment. This sudden change can easily lead to instability at critical points. This invention uses the tanh (hyperbolic tangent) function to map two key physical quantities—the rate of water level change and the water level margin—into a continuously varying allocation ratio α within the interval [0, 1]. When the water level is stable and the margin is large, α approaches 1, and flow is prioritized for power generation. When the water level rises sharply or approaches the flood control limit, α smoothly approaches 0, and flow is continuously and proportionally converted to water abandonment.

[0144] When responding to floods or performing precise intraday scheduling, reservoir water levels may hover near flood control limits or key operating levels. In these situations, the present invention can prevent drastic jumps in power generation output between full and partial output due to minor water level fluctuations. The resulting output power flow curve and abandoned water flow curve are smoother, which not only facilitates stable generator operation and reduces mechanical losses, but also provides a stable and predictable hydropower output curve for the power grid, enhancing the controllability and user-friendliness of hydropower as a high-quality, regulated power source.

Claims

1. A multi-energy complementary collaborative scheduling method based on an improved step-by-step optimization algorithm, characterized in that: include: Obtain hydrological data, engineering data and total grid load; Generate a preliminary feasible dispatch solution used as an initial solution based on hydrological data and total grid load; An improved stepwise optimization algorithm is used to iteratively search for the preliminary feasible scheduling solution to generate the output sequence of each subsystem. Combine the output sequences of each subsystem and redistribute the load according to the deviation from the total load of the power grid to generate an optimized output sequence; Evaluate the optimized output sequence and select the dispatching plan with the greatest peak load-shaving benefit. The optimized output sequence is generated by combining the output sequences of each subsystem and redistributing the load according to the deviation from the total load of the power grid, including: Obtain the current credit value of each reservoir and, combined with the installed capacity in the project data, calculate the credit weighted allocation weight for each reservoir; Determine the total load to be redistributed based on the deviation between the output sequence of each subsystem and the total load of the power grid; Use credit-weighted allocation weights to distribute the total load to be redistributed and generate an optimized output sequence; Calculate the current load rate of each reservoir based on the optimized output sequence and the rated capacity in the project data; Calculate the credit increment based on the deviation between the current load rate and the preset expected load rate, where the credit increment is proportional to the square of the deviation; The credit increment is added to the current credit value of each reservoir to generate an updated credit value for use in the next scheduling cycle.

2. The method according to claim 1, characterized in that Also includes: Maintain a credit value history sequence containing the historical credit values ​​of each reservoir; Add the updated credit value to the credit value history series and calculate the volatility index of the series; When the volatility indicator exceeds the preset volatility threshold, a mean reversion force towards the benchmark credit value is applied to the updated credit value to generate a balanced adjusted credit value, which replaces the updated credit value for use in the next scheduling cycle.

3. The method according to claim 1, characterized in that Based on the hydrological data and the total load of the power grid, a preliminary feasible dispatch solution is generated as the initial solution through an iterative search, which includes: During the iterative search process, the sensitivity of the current hydraulic head to the current power generation flow is calculated; When the water head sensitivity is lower than the preset sensitivity threshold, a virtual potential energy function is constructed to associate the current power generation flow with the target flow; Calculate the gradient of the virtual potential energy function and combine it with the head sensitivity to form a corrected search gradient; The corrected search gradient is used to update the current power generation flow until convergence to obtain a preliminary feasible scheduling solution.

4. The method according to claim 3, characterized in that The virtual potential energy function includes an adaptive elastic coefficient, and its generation method includes: Based on the water head sensitivity, the initial value of the adaptive elastic coefficient is determined, wherein the initial value is inversely proportional to the water head sensitivity exponentially; During the iterative search process, the updated amount of power generation flow is tracked and recorded to form a flow update amount sequence; According to the latest power generation flow update amount and the change trend of the flow update amount sequence, the adaptive elastic coefficient is dynamically attenuated or enhanced; The convergence criteria include: when the changes between the water level and power generation flow calculated in this round of iteration and the corresponding values ​​calculated in the previous round of iteration are both less than their respective preset convergence thresholds, the iteration is judged to have converged.

5. The method according to claim 1, wherein Based on the hydrological data and the total load of the power grid, a preliminary feasible dispatch solution is generated as an initial solution, which further includes the flow allocation process, specifically: Based on the historical water level series and the current water level, calculate the water level change rate at the current moment; Calculate the continuous allocation ratio for dividing the total available flow based on the rate of change of water level; The total available flow is smoothly divided into power generation flow and abandoned water flow using a continuous allocation ratio.

6. The method according to claim 5, characterized in that The steps for calculating the successive allocation ratios for dividing the total available flow include: Calculate the water level margin based on the current water level and the flood control limit water level; The water level change rate and water level margin are used as inputs to generate a continuous allocation ratio through the hyperbolic tangent function.

7. The method according to claim 1, characterized in that The steps for evaluating the optimized output sequence and selecting the dispatching plan with the greatest peak load-shaving benefit include: Establish multi-energy complementary power balance constraints, which are used to measure the consistency between the optimized output sequence and the total load of the power grid; According to the violation degree of power balance constraint, a dynamic penalty function is constructed; The dynamic penalty function is coupled to the scheduling objective function, and the evaluation and selection are completed by solving the scheduling objective function; When the optimized output sequence violates the power balance constraint, the adjustment information used to satisfy the constraint is synchronously fed back to the iterative loop of the improved step-by-step optimization algorithm, and the scheduling core variables within the loop are updated.

8. The method according to claim 1, characterized in that The steps of iteratively searching for a preliminary feasible scheduling solution using the improved stepwise optimization algorithm include: Based on the preliminary feasible scheduling solution and combined with the physical constraints in the engineering data, the dynamic feasible region is delineated; Generate an initial water level process in the dynamic feasible region to serve as the starting point of iterative search; In the dynamic feasible region, the water level at a certain moment in the initial water level process is discretized to generate candidate water levels; The discretization process adopts a variable discrete step size strategy, where the step size is determined according to the current number of iterations and the logarithmic decay law; Distributed search combined with heuristic candidates is used to update decision variables.

Images