Multi-stage scheduling method considering cascade hydropower hierarchical optimization and flexibility constraint

Abstract

The invention discloses a multi-stage scheduling method considering cascade hydropower hierarchical optimization and flexibility constraint, and the method mainly comprises the steps: 1, generating a multi-path uncertainty scene based on a joint probability prediction model of a bidirectional time convolution network-bidirectional long and short term memory-attention mechanism and Gaussian process regression in combination with Latin hypercube sampling; 2, constructing a medium-term scheduling model, and depicting an adjustable section of the system by adjusting structure identification and capability boundary generation; 3, designing a decoupling type multi-target day-ahead scheduling model, and introducing a structure label to identify and adjust a path form; 4, constructing a confidence domain rolling scheduling model, and embedding a season label mechanism to realize dynamic correction; and 5, establishing a cross-stage feedback mechanism, linking middle-stage and day-ahead path adjustment, and unifying a constraint expression structure. The method has the beneficial effects that the adjustment capability, the operation stability and the resource cooperation efficiency of the multi-energy system in an uncertain scene are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of multi-energy complementary optimal scheduling of power systems, and relates to a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower. BACKGROUND

[0002] With the continuous improvement of new energy penetration, power system operation is facing a series of challenges such as deepening of multi-source coupling, increasing of uncertainty, and mismatching of response rhythm, and it is urgent to build an intelligent scheduling mechanism with multi-time scale coordination capability. Although current research has made certain progress in probabilistic prediction, rolling optimization and multi-objective scheduling, there are still many deficiencies in practical application: (1) It is difficult to achieve structural coupling between variables and fine description of probability boundary in wind-solar-load prediction model; (2) The identification method of medium-term regulation resources is relatively static, and lacks dynamic adaptability to future impact sections; (3) The day-ahead multi-objective optimization does not effectively embed the path structure characteristics, and there is a disconnection between scheduling rhythm and execution stability; (4) The day-ahead scheduling lacks response strategies for seasonal load behavior, and it is difficult to cope with the significant energy difference between spring, summer, autumn and winter; (5) There is a lack of systematic feedback mechanism between stages, and the scheduling strategy is difficult to dynamically adjust with operation deviation, which is easy to fall into path dependence and efficiency degradation. SUMMARY

[0003] The purpose of the present application is to provide a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower, in order to reduce the uncertainty of wind power generation and assist the power grid to operate and schedule more effectively.

[0004] The technical solution adopted by the present application is a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower, which is implemented according to the following steps:

[0005] Step 1, an uncertain joint probability prediction method based on bidirectional time convolution neural network-bidirectional long short-term memory network-attention mechanism (BiTCN-BiLSTM-AM) and Gaussian process regression (GPR) is proposed, multivariate time series feature extraction and distribution modeling are used to generate conditional probability output, and Latin hypercube sampling (LHS) is used to construct a multi-path uncertainty scenario;

[0006] Step 2, a dual-module fusion medium-term scheduling method based on regulation structure identification-capability boundary generation is constructed. The regulation structure identification module extracts the time section with potential impact in the wind-solar-load-reservoir response system to form a tunable resource response structure map. The capability boundary generation module sets the upper and lower limit curves of the peak shifting capacity of water power, energy storage and load in combination with the structure map;

[0007] Step 3, a decoupled multi-objective day-ahead scheduling model is constructed, system operation stability, wind and light utilization efficiency and water and electricity regulation capacity are set as target function branches respectively, and a structure label variable is introduced to identify the regulation path form of different schemes;

[0008] Step 4, a confidence region rolling scheduling model is constructed, the deviation range between regulation output and actual response is controlled through the default probability, the load behavior switching mechanism based on the season label is introduced into the model, the load prediction value and the scheduling weight are dynamically corrected, and the strategy rhythm in the short period is reconstructed;

[0009] Step 5, a cross-stage feedback channel is constructed, the operation deviation, water level response and energy storage state change are converted into a feedback matrix, which is used for boundary correction and target adjustment of the medium-term and day-ahead stage, so that the scheduling path is evolved and the structure is converged round by round. At the same time, a unified constraint expression system is established, which integrates power generation balance, hydrological conduction, regulation frequency control and energy storage behavior exclusion logic, so as to realize the solvability of the constraint structure and the consistency of the resource behavior expressed by the hierarchical state variable.

[0010] The characteristics of the application also lie in:

[0011] Step 1 is implemented according to the following implementation:

[0012] The input of the model is a multivariate time series composed of the historical sequence of the corresponding prediction target (wind, light or load) and its related influencing factors, denoted as Wherein represents the input feature of the i-th prediction task at time t, which covers the current variable and its related meteorological influencing factors.

[0013] Firstly, the BiTCN module performs one-dimensional causal convolution operation on X (i) to extract multi-scale local time patterns, and the output is Then, the BiLSTM network further captures the cross-time bidirectional dependency relationship to generate the feature sequence Next, the attention mechanism is used to calculate the attention weight of each time And get the weighted representation Finally, the representation is input into the GPR model, and the conditional distribution of the output prediction value is modeled as:

[0014]

[0015] Wherein, μ(·) is the mean function, and k(·,·) is the kernel function.

[0016] Subsequently, for each variable, the conditional distribution output at each prediction time t is LHS is used to generate multiple uncertainty scenarios. The interval [0, 1] is divided into S equal sub-intervals (i.e. the number of samples is S), and the s-th sampling point is located at:

[0017]

[0018] Finally, the variable transformation is performed by the inverse cumulative distribution function of the Gaussian distribution to obtain sample values that follow the target distribution:

[0019]

[0020] At this point, a multi-scenario path graph is constructed for different variables and different time points, reflecting the uncertain evolution of the future. A multi-path scenario set is formed as input for the subsequent dispatch optimization phase.

[0021] Step 2 is implemented as follows:

[0022] The mid-term scheduling model takes the typical output time series and regulation pressure indicators obtained from the predicted scenario set as input, and combines resource characteristics to perform "regulation structure identification-capacity boundary generation" dual-module scheduling modeling. First, let the system prediction input sequence be:

[0023]

[0024] where, represents the wind and solar power prediction scenario, is the load prediction scenario, and the length of the time series is T.

[0025] (1) Regulation structure identification module:

[0026] By calculating the net load change rate (net load = load - wind and solar power), the system regulation index sequence is constructed:

[0027]

[0028] According to the threshold ω thr , the regulation intensity peak segment is determined, and the regulation structure atlas Ω = {t | ω t ≥ ω thr} is formed, which is used as the structure input for the subsequent capacity boundary.

[0029] (2) Capacity boundary generation module

[0030] According to the regulation structure atlas, combined with the resource characteristics of reservoirs, energy storage, and adjustable loads, the capacity boundary structure of each adjustable resource is set. The hydropower water level regulation capacity is expressed by the following linear boundary:

[0031]

[0032] Wherein, the boundary value is generated by the technical parameters of the corresponding resource and the adjustment demand statistics:

[0033]

[0034] And so on, the upper and lower boundaries of the charging and discharging power of the energy storage system are generated, and the adjustment duration and the minimum continuous operation period requirement T dur .

[0035] Finally, the generated capacity boundary parameter set is represented as:

[0036]

[0037] The set constitutes the boundary input of the adjustable resource capacity and rhythm in the subsequent day-ahead and day-ahead scheduling stage.

[0038] Step 3 is implemented according to the following implementation:

[0039] The day-ahead scheduling model takes the resource capacity boundary set B generated in the medium-term stage and the pre-forecast scenario path As input, constructs a decoupled multi-objective structure around the adjustment capacity, resource utilization efficiency and operation volatility, and introduces a structure label variable for path identification and selection on this basis.

[0040] Let J1, J2, J3 represent the objective function definitions of the maximum adjustment margin, the maximum wind and light absorption efficiency and the minimum output fluctuation respectively, as follows:

[0041]

[0042] Wherein, P hydro , P wind , P pv , P curt represent the water, wind, photovoltaic output and power off power respectively; P t total =P t hydro +P t wind +P t pv .

[0043] Introducing a structure label variable σ s,t ∈{0,1} T , used to represent the adjustment section distribution characteristics of the s-th path in the time domain, and is specifically defined as:

[0044]

[0045] For each candidate path s∈[1,S], calculate its structure label vector σ sThe target performance and the regulation structure information are integrated by a weighted vector scoring function:

[0046]

[0047] where is the normalized value of the objective function J i under the s-th path, std(σ s ) represents the standard deviation of the regulation frequency of the path (measuring the fluctuation of the regulation rhythm), and the weight coefficients α i , β are set according to the system scheduling strategy.

[0048] Finally, the solution that minimizes Ψ s is selected from all candidate paths as the reference path for day-ahead scheduling:

[0049] s * = argmin s∈[1,S] Ψ s (16)

[0050] Step 4 is implemented as follows:

[0051] Based on the day-ahead scheduling reference path , the system scheduling strategy is modified in each rolling period τ ∈ [1, T roll ]. This model combines the prediction confidence interval control and the seasonal behavior label switching mechanism to adjust the load structure and resource invocation weight while meeting the scheduling safety boundary, achieving rolling adaptive optimization in short periods.

[0052] (1) Confidence domain control mechanism

[0053] Let the predicted value of the i-th input variable (such as wind speed, load) in the rolling period τ follow a conditional Gaussian distribution:

[0054]

[0055] Define its upper and lower confidence limits as:

[0056]

[0057] where z α is the quantile coefficient of the normal distribution corresponding to the confidence level α, and when z α ≈ 1.96.

[0058] In rolling scheduling, the system output P τ total must satisfy:

[0059]

[0060] (2) Confidence domain control mechanism

[0061] Let the current season label be θ∈{spring, summer, autumn, winter}, and define the corresponding load correction factor function according to each type of label:

[0062]

[0063] Apply the correction factor to the rolling load prediction:

[0064]

[0065] The updated load prediction result is used as the target output reference to participate in the rolling optimization scheduling model, which is used to dynamically adjust the priority and output target of resources such as hydropower, energy storage, wind and light.

[0066] Step 5 is implemented according to the following:

[0067] In the dispatching system operation cycle, in order to improve the cross-stage coordination and adaptive ability, an information feedback mechanism is built with “operation deviation feedback” as the core, covering the medium-term, day-ahead and intra-day three stages. At the same time, a unified constraint expression framework is established to ensure that the dispatching results meet the system operation boundary and feasibility requirements.

[0068] (1) Operation deviation feedback mechanism

[0069] Let the actual output of the system be P t actual , and the target path output be P t plan , then the scheduling deviation can be defined as:

[0070] ε t = P t actual - P t plan (23)

[0071] The cumulative deviation reflects the deviation trend of the system execution trajectory from the scheduling path:

[0072]

[0073] To guide the path update, introduce the deviation feedback weight matrix Modify the objective function or constraint boundary:

[0074]

[0075] where, represents the kth feedback corrected objective function component, λ k is the feedback gain coefficient.

[0076] (2) Unified constraint expression structure

[0077] To keep the model solution space closed and consistent with the resource behavior, a unified constraint system C is constructed, covering the following substructures:

[0078] Power balance constraint:

[0079] P t hydro +P t wind +P t pv +P t dis =P t load +P t ch +P t curt (26)

[0080] Hydrological linkage and water level change constraint:

[0081]

[0082] Energy storage charging and discharging exclusion constraint (logical type):

[0083] P t ch ·P t dis =0 (28)

[0084] Hydropower regulation wave control constraint:

[0085]

[0086] Section transmission capacity constraint (node j):

[0087]

[0088] Finally, all the above constraints are combined into the feasible region of the model solution. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 is a multi-time scale collaborative scheduling general framework in a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower;

[0090] Figure 2 is a BiTCN-BiLSTM-AM-GPR joint prediction structure diagram in a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower;

[0091] Figure 3is a medium-term regulation structure-border response graph in a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower;

[0092] Figure 4 is a multi-path structure label scoring and screening flow chart in a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower;

[0093] Figure 5 is a rolling feedback mechanism and unified constraint system diagram in a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower;

[0094] Figure 6 is a multi-season wind speed and solar radiation prediction and scenario simulation diagram in a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower: (a) wind speed, (b) solar radiation;

[0095] Figure 7 is a cascade hydropower optimization scheduling result diagram in a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower in a typical spring day. DETAILED DESCRIPTION

[0096] The present application will be described in detail below in conjunction with the drawings and specific embodiments.

[0097] As shown in Figure 1 , the present application as a whole adopts a multi-time scale collaborative scheduling framework, covering joint prediction, medium-term evaluation, day-ahead optimization, intra-day correction and closed-loop feedback, forming a scheduling chain that is progressive and dynamically adjusted.

[0098] The present application is a multi-stage scheduling method considering hierarchical optimization and flexibility constraints of cascade hydropower, which is implemented according to the following steps:

[0099] Step 1 is implemented according to the following:

[0100] The input of the model is a multivariate time series composed of the historical sequence of the corresponding prediction target (wind, light or load) and its related influencing factors, denoted as where represents the input feature of the i-th prediction task at time t, covering the current variable and its related meteorological influencing factors.

[0101] First, the BiTCN module performs one-dimensional causal convolution operation on X (i) to extract multi-scale local time patterns, and the output is Then, the BiLSTM network further captures the cross-time bidirectional dependency relationship to generate the feature sequence Next, the attention mechanism is used to calculate the attention weight of each time and obtain the weighted representation Finally, the characterization is input into the GPR model, modeling the conditional distribution of the output prediction value as:

[0102]

[0103] where μ(·) is the mean function and k(·,·) is the kernel function.

[0104] Subsequently, the conditional distribution of each variable output at each prediction time t is LHS is used to generate multiple uncertainty scenarios. The interval [0, 1] is divided into S equal subintervals (i.e., the number of samples is S), and the s-th sampling point is located at:

[0105]

[0106] Finally, the inverse cumulative distribution function of the Gaussian distribution is used for variable transformation to obtain sample values subject to the target distribution:

[0107]

[0108] At this point, multiple scenario path graphs for different variables and different time points are constructed, reflecting the uncertain evolution in the future. A multi-path scenario set is formed as input for the subsequent scheduling optimization stage. The joint prediction overall structure process is as shown in Figure 2 .

[0109] Step 2 is implemented according to the following:

[0110] The mid-term scheduling model takes the typical output time sequence and regulation pressure index obtained from the prediction scenario set as input, and combines resource characteristics to perform "regulation structure identification-capability boundary generation" dual-module scheduling modeling. First, let the system prediction input sequence be:

[0111]

[0112] wherein represents the wind and solar power prediction scenario, is the load prediction scenario, and the length of the time sequence is T.

[0113] (1) Regulation structure identification module:

[0114] By calculating the net load change rate (net load = load - wind and solar power), the system regulation index sequence is constructed:

[0115]

[0116] According to the threshold ω thr , the regulation intensity peak segment is determined, and the regulation structure map Ω = {t | ω t ≥ ωthr}, the structure input for subsequent capacity boundary.

[0117] (2) Capacity boundary generation module

[0118] According to the regulation structure map, combined with the resource characteristics of reservoirs, energy storage and adjustable loads, the capacity boundary structure of each adjustable resource is set. The regulation capacity of hydropower is expressed by the following linear boundary:

[0119]

[0120] Wherein, the boundary value is generated by the technical parameters of the corresponding resource and the regulation demand statistics:

[0121]

[0122] And so on, the charge and discharge power upper and lower boundaries of the energy storage system are generated, and the regulation time length and the minimum continuous operation period requirement T dur .

[0123] Finally, the generated capacity boundary parameter set is represented as:

[0124]

[0125] The set constitutes the boundary input of adjustable resource capacity and rhythm in subsequent day-ahead and intra-day dispatching stages. As Figure 3 shown, the water level curve is analyzed according to the incoming water process in the early stage of dispatching, the high pressure boundary and the lower limit of the reservoir before dispatching are identified through the regulation structure, the dispatching reserve capacity area is demarcated, and the capacity boundary interval is constructed combined with the feedback regulation mechanism, providing boundary support for the subsequent dispatching stage.

[0126] Step 3 is implemented according to the following:

[0127] The day-ahead dispatching model takes the resource capacity boundary set B generated in the medium-term stage and the pre-forecast scenario path as input, constructs a decoupled multi-objective structure around regulation capacity, resource utilization efficiency and operation volatility, and introduces a structure label variable for path identification and selection on this basis.

[0128] Let J1, J2, J3 represent the objective function definitions of regulation margin maximization, wind and light absorption efficiency maximization and output fluctuation minimization respectively as follows:

[0129]

[0130] Wherein, P hydro , P wind , P pv , P curt represent the output of hydropower, wind power, photovoltaic and abandoned power respectively; Pt total =P t hydro +P t wind +P t pv .

[0131] Introducing the structural label variable σ s,t ∈{0,1} T , which is used to represent the distribution characteristics of the regulation section of the sth path in the time domain, is specifically defined as:

[0132]

[0133] For each candidate path s∈[1,S], calculate its structural label vector σ s , and then the weighted vector scoring function is used to integrate the target performance and the adjustment structure information:

[0134]

[0135] in is the objective function J under the sth path i The normalized value, std(σ s ) represents the standard deviation of the adjustment frequency of the path (measures the fluctuation of the adjustment rhythm), and the weight coefficient α i ,β is set according to the system scheduling strategy.

[0136] Finally, choose Ψ among all candidate paths s The minimum solution is used as the reference path for day-ahead scheduling:

[0137] s * =argmin s∈[1,S] Ψ s (16)

[0138] like Figure 4 As shown in the figure, the path screening process dynamically evaluates candidate paths based on multiple indicators such as structural integrity, score threshold, and goal achievement, gradually eliminates paths that do not meet the requirements and adjusts label priorities, and finally outputs the optimal path label set that meets multi-objective scheduling requirements.

[0139] Step 4 is implemented as follows:

[0140] Day-ahead scheduling reference path Based on each rolling period τ∈[1,T roll ] to modify the system scheduling strategy within a certain period. This model combines prediction confidence interval control with seasonal behavior label switching mechanism to adjust the load structure and resource call weight while meeting the scheduling safety margin, achieving rolling adaptive optimization within a short period.

[0141] (1) Confidence domain control mechanism

[0142] Let the prediction value of the ith input variable (such as wind speed, load) in the rolling period τ obey the conditional Gaussian distribution:

[0143]

[0144] Define its upper and lower confidence limits as:

[0145]

[0146] where z α is the quantile coefficient corresponding to the confidence level α under the normal distribution, and when z α ≈1.96.

[0147] The system output in the rolling dispatch is required to satisfy:

[0148]

[0149] (2) Confidence domain control mechanism

[0150] Let the current seasonal label be θ ∈ {spring, summer, autumn, winter}, and define the corresponding load correction factor function according to each type of label:

[0151]

[0152] Apply the correction factor to the rolling load prediction:

[0153]

[0154] The updated load prediction result is used as the target output benchmark to participate in the rolling optimization dispatch model, and is used to dynamically adjust the priority and output target of resources such as hydropower, energy storage, wind and solar.

[0155] Step 5 is implemented according to the following:

[0156] In the dispatch system operation cycle, in order to improve the cross-stage coordination and adaptive ability, an information feedback mechanism is constructed with "operation deviation feedback" as the core, covering the medium-term, day-ahead and intra-day three stages. At the same time, a unified constraint expression framework is established to ensure that the dispatch results meet the system operation boundary and feasibility requirements.

[0157] (1) Operation deviation feedback mechanism

[0158] Let the actual output of the system be P t actual , and the target path output be P t plan ​The scheduling bias can be defined as:

[0159] ε t = P t actual - P t plan (23)

[0160] The cumulative bias reflects the trend of the system trajectory deviating from the scheduling path:

[0161]

[0162] To guide path updates, introduce bias feedback weight matrix Modify the objective function or constraint boundary:

[0163]

[0164] where, represents the kth feedback modified objective function component, λ k is the feedback gain coefficient.

[0165] (2) Unified constraint expression structure

[0166] To maintain the closure of the model solution space and the consistency of resource behavior, a unified constraint system C is constructed, which covers the following substructures:

[0167] Power balance constraint:

[0168] P t hydro + P t wind + P t pv + P t dis = P t load + P t ch + P t curt (26)

[0169] Hydrological linkage and water level change constraint:

[0170]

[0171] Energy storage charging and discharging exclusion constraint (logical type):

[0172] P t ch · P t dis = 0 (28)

[0173] Hydropower regulation wave control constraint:

[0174]

[0175] Section transmission capacity constraint (node ​​j):

[0176]

[0177] Finally, all the above constraints are combined into the feasible region of the model solution.

[0178] like Figure 5 As shown in the figure, the rolling feedback input serves as the key fulcrum of rolling control, pointing to the rolling feedback correction, and its structure simultaneously accepts two information flows from the global constraint dynamic correction and the disturbance prediction feedback, realizing a closed-loop mechanism of disturbance perception-path correction-structure adjustment, effectively enhancing the adaptive ability of the system.

[0179] Example

[0180] Taking the wind speed and solar radiation data of four typical seasons in spring, summer, autumn and winter in Sichuan Province in 2021 as an example, the prediction accuracy of renewable resources and the representativeness of simulation scenarios are analyzed. Figure 6 Figures (a) and (b) show the seasonal predictions and 95% confidence intervals for wind speed and solar irradiance, respectively, along with the various simulated scenarios constructed. Wind speed fluctuates somewhat across seasons, with higher wind speeds in summer and autumn. Solar irradiance exhibits a typical diurnal cycle, with significant differences between the seasons. The simulated scenarios generally cover the range of observed values, demonstrating high credibility and representativeness.

[0181] The cascaded multi-energy complementary system constructed in this invention includes three hydropower stations and their grid-connected wind and photovoltaic power generation systems. The wind farm and photovoltaic power station each have an installed capacity of 500 MW and are connected to the second and first hydropower stations, respectively. The regulation functions of the units within each hydropower station are divided as shown in Table 1.

[0182] Table 1 Parameters related to cascade hydropower configuration

[0183]

[0184] On this basis, a typical spring day was selected as the scheduling input to carry out the optimization scheduling experiment of the wind, solar, hydropower and storage multi-energy complementary system. Figure 7 The dispatch results of the three strategies under spring conditions are presented. It can be seen that, while ensuring stable system output, the scheme effectively utilizes the hydropower regulation capacity, fully accommodates renewable energy output, significantly reduces curtailment, achieves excellent load tracking, and achieves ideal dispatch response, validating the effectiveness and adaptability of the proposed model and strategy.

Claims

1. A multi-level scheduling method considering cascade hydropower hierarchical optimization and flexibility constraints, characterized by: Please follow the steps below to implement it: Step 1: Propose an uncertain joint probability prediction method based on a bidirectional temporal convolutional neural network-bidirectional long short-term memory network-attention mechanism (BiTCN-BiLSTM-AM) and Gaussian process regression (GPR). This method uses multivariate time series feature extraction and distribution modeling to generate conditional probability outputs, and combines Latin hypercube sampling (LHS) to construct multipath uncertainty scenarios. Step 2: Construct a dual-module fusion medium-term scheduling method based on regulation structure identification and capacity boundary generation. The regulation structure identification module extracts time segments with potential impacts in the wind-solar-load-reservoir response system and forms a structural map of adjustable resource response. The capacity boundary generation module combines this structural map to set upper and lower limit curves for hydropower, energy storage, and load peak shifting capabilities. Step 3: Construct a decoupled multi-objective day-ahead dispatch model, setting system operation stability, wind and solar utilization efficiency, and hydropower regulation capacity as objective function branches respectively, and introduce structural label variables to identify the regulation path forms of different schemes; Step 4: Build a confidence region rolling dispatch model, using the day-ahead path as a structural reference. This model uses default probability control to adjust the deviation between output and actual response. The model also incorporates a seasonally labeled load behavior switching mechanism, dynamically revising load forecasts and dispatch weights to reconstruct the strategic rhythm within a short period. Step 5: Construct a cross-stage feedback channel, transforming operational deviations, water level responses, and energy storage state changes into a feedback matrix. This matrix acts on boundary corrections and target adjustments during the mid-term and day-ahead phases, achieving round-by-round evolution and structural convergence of the dispatch path. Simultaneously, a unified constraint expression system is established, integrating the mutually exclusive logic of power generation balance, hydrological transmission, regulation frequency control, and energy storage behavior. This system expresses resource relationships using hierarchical state variables, ensuring the solvability of the constraint structure and the coordinated consistency of resource behavior.

2. A multi-level scheduling method considering cascade hydropower layer optimization and flexibility constraints according to claim 1, characterized in that: The step 1 is specifically implemented as follows: The input of the model is the multivariate time series consisting of the historical sequence of the corresponding forecast target (wind, light or load) and its related influencing factors, which is recorded as in It represents the input features of the i-th prediction task at time t, covering the current variables and their related meteorological influencing factors. First, the BiTCN module (i) A one-dimensional causal convolution operation is performed to extract multi-scale local temporal patterns, and the output is Then, the BiLSTM network further captures the bidirectional dependencies across time periods and generates feature sequences Next, the attention weight of each moment is calculated through the attention mechanism And obtain the weighted representation Finally, the representation is input into the GPR model, and the conditional distribution of the model output prediction value is: Among them, μ(·) is the mean function and k(·,·) is the kernel function. Then, the conditional distribution output for each variable at each prediction time t is Use LHS to generate multiple uncertainty scenarios. Divide the interval [0,1] into S equal-width subintervals (i.e., the number of samples is S), and the sth sampling point is located at: Finally, the variable is transformed by the inverse cumulative distribution function of the Gaussian distribution to obtain the sample value that obeys the target distribution: At this point, a multi-scenario path diagram with different variables and at different times is constructed to reflect the uncertain evolution of the future. Serves as input for the subsequent scheduling optimization phase.

3. A multi-level scheduling method considering cascade hydropower layer optimization and flexibility constraints according to claim 1, characterized in that: The step 2 is specifically implemented as follows: Step 2 is implemented as follows: The medium-term scheduling model uses the typical output time series and regulation pressure indicators obtained from the forecast scenario set as input, and combines resource characteristics to perform a dual-module scheduling model of "regulation structure identification and capacity boundary generation". First, assume that the system forecast input sequence is: in, Indicates the wind and solar power output forecast scenario, It is a time series with a length of T for load forecasting scenario. (1) Adjusting the structure recognition module: By calculating the net load change rate (net load = load - wind and solar output), a system regulation index sequence is constructed: According to the threshold ω thr Determine the peak segment of the regulation intensity and form the regulation structure map Ω={t|ω t ≥ω thr }, used for structural input of subsequent capability boundaries. (2) Capability boundary generation module Based on the regulation structure map, combined with the resource characteristics of reservoirs, energy storage, and adjustable loads, the capacity boundary structure of each adjustable resource is set. The hydropower water level regulation capacity is expressed as follows: The boundary value is generated by the technical parameters of the corresponding resource and the adjustment demand statistics: And so on, generate the upper and lower limits of charge and discharge power for the energy storage system, adjust the duration and the minimum continuous operation period requirement T dur . Finally, the capacity boundary parameter set generated above is expressed as: This set constitutes the boundary input for the adjustable resource capacity and rhythm in the subsequent day-ahead and intraday scheduling stages.

4. A multi-level scheduling method considering cascade hydropower hierarchical optimization and flexibility constraints according to claim 1, characterized in that: The step 3 is specifically implemented as follows: Step 3 is implemented as follows: The day-ahead scheduling model uses the resource capacity boundary set B generated in the mid-term phase and the pre-forecast scenario path Taking the regulatory capacity, resource utilization efficiency and operational volatility as input, a decoupled multi-objective structure is constructed, and on this basis, structural label variables are introduced for path identification and selection. Assume that J1, J2, and J3 represent the objective functions of maximizing the regulation margin, maximizing the wind and solar absorption efficiency, and minimizing the output volatility, respectively. The definitions are as follows: Among them, P hydro 、P wind 、Ppv、P curt Respectively represent hydropower, wind power, photovoltaic power output and abandoned power; Introducing the structural label variable σ s,t ∈{0,1} T , which is used to represent the distribution characteristics of the regulation section of the sth path in the time domain, is specifically defined as: For each candidate path s∈[1,S], calculate its structural label vector σ s , and then the weighted vector scoring function is used to integrate the target performance and the adjustment structure information: in is the objective function J under the sth path i The normalized value, std(σ s ) represents the standard deviation of the adjustment frequency of the path (measures the fluctuation of the adjustment rhythm), and the weight coefficient α i ,β is set according to the system scheduling strategy. Finally, choose Ψ among all candidate paths s The minimum solution is used as the reference path for day-ahead scheduling: s * =argmin s∈[1,S] P s (16)。 5. The multi-level scheduling method according to claim 1, which takes into account the hierarchical optimization and flexibility constraints of cascade hydropower, is characterized in that: The step 4 is specifically implemented as follows: Step 4 is implemented as follows: Day-ahead scheduling reference path Based on each rolling period τ∈[1,T roll ] to modify the system scheduling strategy within a certain period. This model combines prediction confidence interval control with seasonal behavior label switching mechanism to adjust the load structure and resource call weight while meeting the scheduling safety margin, achieving rolling adaptive optimization within a short period. (1) Trust region control mechanism Assume that the predicted value of the i-th input variable (such as wind speed, load) in the rolling period τ obeys the conditional Gaussian distribution: Define its confidence upper and lower bounds as: Among them, z α is the quantile coefficient corresponding to the confidence level α under normal distribution, and z is taken as 0.95 α ≈1.

96. System output required in rolling scheduling satisfy: (2) Trust region control mechanism Assume that the current season label is θ∈{spring, summer, autumn, winter}, and define the corresponding load correction factor function according to each type of label: Apply the correction factor to the rolling load forecast: The updated load forecast results are used as the target output benchmark in the rolling optimization scheduling model to dynamically adjust the priority and output targets of hydropower, energy storage, wind and solar resources.

6. A multi-level scheduling method considering cascade hydropower layer optimization and flexibility constraints according to claim 1, characterized in that: The step 5 is specifically implemented as follows: Step 5 is implemented as follows: To enhance cross-stage coordination and adaptability within the dispatch system's operational cycle, an information feedback mechanism centered around operational deviation feedback was established, covering the mid-term, day-ahead, and intraday phases. Furthermore, a unified constraint expression framework was established to ensure dispatch results met the system's operational boundaries and feasibility requirements. (1) Operational deviation feedback mechanism Assume that the actual output of the system is The target path output is Then the scheduling deviation can be defined as: The cumulative deviation reflects the deviation trend of the system execution trajectory from the scheduling path: To guide path updates, a bias feedback weight matrix is ​​introduced Modify the objective function or constraint bounds: in, represents the objective function component after the kth feedback correction, λ k is the feedback gain coefficient. (2) Unified constraint expression structure To maintain the closure of the model solution space and the consistency of resource behavior, a unified constraint system C is constructed, covering the following substructures: Power balance constraints: Hydrological linkage and water level change constraints: Energy storage charging and discharging mutual exclusion constraints (logical type): Hydropower regulation and wave control constraints: Section transmission capacity constraint (node ​​j): Finally, all the above constraints are combined into the feasible region of the model solution.

Images