An adaptive switching method for intelligent scheduling of an energy system

Abstract

This invention relates to the field of circuit devices or systems for power supply or distribution, and discloses an adaptive switching method for intelligent scheduling of energy systems. The method includes: initializing sub-models, including an artificial neural network proxy model and a running priority model, as scheduling candidates; obtaining net load forecasts based on forecast data, calculating forecast uncertainty indicators, net load indicators, and risk indicators; calculating a comprehensive risk index by weighted summation of all indicators, comparing it with a preset threshold, and selecting a high-risk model if the comprehensive risk index is higher than the threshold; otherwise, selecting a non-high-risk model; generating a scheduling scheme based on the selected model, and verifying the feasibility of the scheme; if operational constraints are violated, switching to other model schemes and re-verifying. This method solves the problems of reliance on a single algorithm, computational delay, insufficient optimization, and low security in existing technologies, achieving the goals of dynamic scheduling, economic efficiency, security, reliability, and strong scalability.

Description

Technical Field

[0001] This invention relates to the field of circuit devices or systems for power supply or distribution, and in particular to an adaptive switching method for intelligent scheduling of energy systems. Background Technology

[0002] In existing technologies, energy system dispatching mainly relies on a single algorithm: methods based on mathematical optimization (such as mixed-integer linear programming) can achieve better economic results, but their high computational complexity makes it difficult to meet the requirements of real-time and rapid solution; methods based on rule or priority logic can quickly output dispatching schemes and strictly meet operational constraints, but they lack global optimization capabilities, easily leading to increased wind and solar curtailment and higher operating costs; methods based on surrogate models such as neural networks can output approximate solutions in milliseconds, significantly reducing computational latency, but may produce infeasible solutions that do not meet power boundaries or power balance under conditions of large prediction deviations or off-grid operation. Therefore, existing technologies generally suffer from the inability to simultaneously achieve real-time performance, economy, and security. This paper proposes a dispatching algorithm switching method that enables energy systems to automatically switch between multiple dispatching algorithms based on different operating scenarios: when the system is grid-connected, operating stably, and predictions are accurate, a surrogate model is used as the primary method to achieve fast and economical dispatching; when the system is isolated, experiencing severe fluctuations, or predictions are inaccurate, a rule-based dispatching method is switched to ensure that the operating results strictly meet constraints such as power balance and output boundaries. This mechanism allows for the full utilization of the advantages of different algorithms, ensuring both the real-time performance and economy of scheduling, while also guaranteeing the safety and reliability of system operation under critical conditions.

[0003] For example, Chinese patent CN117810989A discloses a method for establishing and solving a data mechanism-driven power grid optimal scheduling model. It provides the following technical solution: establishing a physical model of a comprehensive energy system with uncertainties injected; generating simulation cases based on the physical model, and generating corresponding solution data as a training set for machine learning; training the machine learning model for learning and solving; integrating the physical model with the data model to establish a data mechanism-driven power grid optimal scheduling model, and proposing a data mechanism-driven optimal scheduling method for power grid systems. This improves the accuracy of deep neural network mapping, significantly reduces scheduling solution time through the data-driven approach, and achieves coordinated optimization of optimal renewable energy carrying capacity and controllable equipment planning and operation, thereby improving renewable energy utilization.

[0004] However, the aforementioned method for establishing and solving a data-mechanism hybrid-driven power grid optimization scheduling model emphasizes the integration of mechanism models and data-driven models. It establishes a physical optimization model of the integrated energy system, then uses Gurobi to generate training data, trains a Transformer-based deep neural network, and finally couples the mechanism model and data model to form a unified optimization scheduling framework driven by a "data-mechanism hybrid approach." Its core is to achieve the integration of mechanisms and data within a unified model, thereby improving the model's prediction accuracy and scheduling efficiency. This patent's core remains integration within a unified model, lacking the ability to differentiate scheduling methods under different operating conditions. When the system is in a scenario of severe fluctuations or large prediction deviations, the model's adaptability may still be insufficient. This invention, however, proposes an algorithm switching mechanism. Instead of a single model fusion, it dynamically switches between different scheduling algorithms based on the uncertainty of wind and solar forecasts and the system's operating state, thereby integrating the advantages of different scheduling optimization models. Summary of the Invention

[0005] This invention addresses the problems of reliance on a single algorithm, computational latency, insufficient optimization, and low security in existing technologies. It proposes an adaptive switching method for intelligent energy system scheduling, achieving dynamic scheduling, economic efficiency, security, reliability, and strong scalability. By real-time evaluation of risk indicators such as wind and solar forecast uncertainties and net load fluctuations, the system prioritizes the use of a neural network surrogate model under stable operating conditions to achieve millisecond-level response and cost optimization. Under fluctuating or extreme conditions, it automatically switches to a rule-priority algorithm to ensure strict compliance with operational constraints. This fully leverages the complementary advantages of different algorithms, improving overall scheduling efficiency and stability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An adaptive switching method for intelligent scheduling of an energy system includes:

[0008] Initialize sub-models, including artificial neural network agent models and running priority models, as scheduling candidates;

[0009] Based on the forecast data, the net load forecast value is obtained, and the forecast uncertainty index, net load index, and risk index are calculated.

[0010] Calculate the comprehensive risk index by weighted summation of all indicators, and compare it with a preset threshold. If the comprehensive risk index is higher than the threshold, select the high-risk model; otherwise, select the non-high-risk model.

[0011] A scheduling scheme is generated based on the selected model, and the feasibility of the model scheme is verified. If the operational constraints are violated, the scheme is switched to another model scheme and verified again.

[0012] By initializing multiple sub-models as scheduling candidates, the system can dynamically switch algorithms according to the operating scenario, thereby pursuing economy under stable conditions and ensuring safety under fluctuating or extreme conditions. This solves the problem that a single algorithm cannot simultaneously achieve real-time performance, economy, and safety, fully leveraging the complementary advantages of different algorithms to improve overall scheduling efficiency.

[0013] Preferably, the calculation of the prediction uncertainty index specifically includes: obtaining the wind and solar power output prediction range for the future target time step, which is obtained by subtracting the lower limit value from the upper limit value of the wind and solar power output prediction; dividing the wind and solar power output prediction range by the wind and solar installed capacity in the energy system to obtain a normalized prediction uncertainty index; the prediction uncertainty index quantifies the accuracy of wind and solar power prediction, and the larger the index value, the higher the prediction uncertainty, and the less suitable it is to use an artificial neural network surrogate model.

[0014] Normalization makes the forecast uncertainty indicators comparable, enabling accurate quantification of the volatility of wind and solar forecasts. Larger indicator values ​​make surrogate models less suitable, which helps to automatically switch to a safer scheduling method when forecasts are inaccurate, avoiding infeasible solutions due to forecast bias and improving system reliability.

[0015] Preferably, the calculation of the net load ramp-up risk index specifically includes: assessing the net load change rate at the future target time step, which is obtained by subtracting the net load value at the future first time step from the net load value predicted at the current moment for the second future time step, taking the absolute value, and then dividing by the time interval; then comparing the net load change rate with the system's allowable adjustment rate, which is calibrated by the comprehensive adjustment capability of non-renewable energy generator sets and energy storage devices; the closer the net load ramp-up risk index value is to 1, the closer the net load fluctuation rate is to the system's safety limit, and the less suitable it is to use an artificial neural network proxy model.

[0016] By comparing the net load change rate with the system's safe adjustment rate, this indicator can identify ramp-up risks in advance. When the indicator value approaches 1, it switches to a priority model to ensure that the system can still meet power balance and output boundary constraints during rapid load fluctuations, preventing safety incidents from occurring.

[0017] Preferably, the power shortage risk index and the power curtailment risk index specifically include: For the power shortage risk index, when the predicted wind and solar power output is low, the estimated net load calculated using the lower limit of the predicted wind and solar power output under pessimistic conditions is used. If this value exceeds the preset proportional threshold of the upper limit of the output of non-renewable energy regulating units, then a power shortage risk is determined, and the index is marked as 1; otherwise, it is marked as 0. For the power curtailment risk index, when the predicted wind and solar power output is high, the estimated net load calculated using the upper limit of the predicted wind and solar power output under optimistic conditions is used. If this value is lower than the preset proportional threshold of the lower limit of the output of non-renewable energy regulating units, then a power curtailment risk is determined, and the index is marked as 1; otherwise, it is marked as 0.

[0018] The risk identification process has been simplified, and indicators for power shortages and curtailment risks can provide rapid early warnings of power shortages or renewable energy waste. This helps to switch scheduling algorithms in a timely manner under extreme conditions, reduce wind and solar curtailment, ensure power supply continuity, and optimize resource utilization.

[0019] Preferably, the net load volatility index specifically includes: calculating the square of the difference between the net load forecast values ​​at adjacent time points, and normalizing it based on the installed capacity of renewable energy; the larger the volatility index value, the more drastic the future net load changes, the higher the scheduling difficulty, and the more suitable it is to prioritize the operation priority model.

[0020] Normalization makes the volatility index applicable to systems of different sizes; the larger the index value, the higher the priority model is applied, ensuring that the scheduling scheme strictly meets constraints when the net load changes drastically. This improves the algorithm's adaptability to complex operating conditions and reduces the risk of scheduling failure.

[0021] Preferably, when calculating the comprehensive risk index using weighted summation, the weight values ​​are determined through offline calibration and optimization; the preset threshold includes an upper threshold, which is optimized through historical data simulation with the goal of minimizing the annual operating cost; the comprehensive risk index is the sum of each indicator value multiplied by its corresponding weight, and the comparison between the index and the threshold distinguishes between high-risk and low-risk scenarios.

[0022] Weights and thresholds are optimized and calibrated using historical data, ensuring that the comprehensive risk index accurately reflects the system's operational risks. Minimizing operating costs is the goal, ensuring a balance between economy and security in the algorithm switching mechanism and improving long-term operational efficiency.

[0023] Preferably, the feasibility verification specifically includes: when an artificial neural network proxy model is selected, checking whether the scheduling scheme generated by it meets the system operation constraints, including the upper and lower limits of unit output, the energy storage regulation boundary, and the power balance constraints; if any constraint is violated, the scheduling scheme generated by the operation priority model is switched to.

[0024] Feasibility verification ensures that the scheduling scheme output by the proxy model is feasible in actual operation. If any constraint is violated, the system immediately switches to the priority model to avoid system failures caused by infeasible solutions and enhances the stability of the scheduling results.

[0025] Preferably, the weight value is optimized through offline calibration, which includes: simulating the scheduling process and calculating the operating cost at each time step based on historical full-year data. The operating cost includes fossil fuel costs, equipment maintenance costs, and power curtailment penalty costs. With the goal of minimizing the total annual operating cost, the combination of weight and threshold parameters is optimized by enumeration or search, and the parameter with the lowest total cost is selected as the set value.

[0026] Offline calibration is based on simulations using historical full-year data, comprehensively considering fuel costs, operation and maintenance costs, and curtailment penalties to ensure scientifically sound parameter settings. By minimizing total operating costs, the algorithm switching mechanism is guaranteed to be economical and efficient in real-world scenarios.

[0027] Preferably, the non-high-risk model adopts an artificial neural network proxy model, and the high-risk model adopts a running priority model; the switching to other models specifically includes subdividing several models by risk index range, and when the comprehensive risk index is in different ranges, it corresponds to different scheduling models, including robust optimization models and heuristic search models.

[0028] The framework supports algorithm expansion, and the subdivision of the risk index range allows for the integration of more algorithms such as robust optimization or heuristic search. This enhances its versatility and evolvability.

[0029] Preferably, the net load index includes a net load ramp-up risk index and a net load volatility index; the risk index includes a power shortage risk index and a power curtailment risk index; the non-high-risk model adopts an artificial neural network surrogate model, and the high-risk model adopts an operation priority model; the prediction data includes wind and solar power output probability prediction data and load prediction data for future times.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0031] 1. This invention introduces an algorithm switching mechanism driven by predictive data. The system can assess the uncertainty indicators and net load fluctuations of wind and solar power forecasts in real time, thereby intelligently selecting the most suitable scheduling algorithm. In scenarios with stable operation and accurate predictions, an artificial neural network surrogate model is prioritized to improve scheduling efficiency and reduce computational latency. However, when the system faces drastic fluctuations in wind and solar power output or increased prediction deviations, it automatically switches to a rule-based algorithm with a set priority, ensuring that the scheduling scheme strictly meets power balance and equipment constraints, and avoiding the generation of infeasible solutions. This dynamic switching capability frees the energy system from the limitations of a single algorithm, allowing for flexible adjustments based on actual operating conditions. This ensures real-time response speed and enhances the system's adaptability to complex operating conditions, ultimately achieving intelligent and precise scheduling.

[0032] 2. This invention effectively solves the problem of traditional scheduling methods failing to balance optimization objectives and operational constraints through a comprehensive risk index assessment mechanism. Under stable operating conditions, the system prioritizes economic efficiency, employing a proxy model to quickly generate scheduling schemes, significantly reducing fossil fuel consumption and operation and maintenance costs, while also minimizing wind and solar power curtailment. Under fluctuating or extreme conditions, the system automatically switches to a safety-oriented priority algorithm to ensure timely response from energy storage and standby units, preventing power shortages or equipment overload. This scenario-based optimization strategy not only improves overall operational efficiency but also minimizes operating costs through offline calibration of weights and thresholds and simulation using historical data, enabling the system to maintain low energy consumption and high reliability throughout long-term operation, achieving a synergistic improvement in both economic efficiency and safety.

[0033] 3. The algorithm switching architecture proposed in this invention possesses high versatility and evolution potential, and is not limited to specific algorithm types. Through the core mechanism of risk index and threshold determination, the system can easily integrate multiple scheduling methods, such as robust optimization or heuristic search, and dynamic switching between multiple algorithms can be achieved simply by adjusting the corresponding relationship of risk intervals. This design allows energy systems to continuously introduce new algorithms as technology advances without reconstructing the overall framework, reducing upgrade costs and enhancing the system's future adaptability. Furthermore, the decoupling characteristics of the framework allow for customized parameters based on different energy scales, and threshold settings can be optimized through historical simulations to ensure that the extended algorithms remain efficient and robust. Attached Figure Description

[0034] Figure 1 This is an overall flowchart of an adaptive switching method for intelligent scheduling of an energy system according to the present invention.

[0035] Figure 2 This is a typical monthly system scheduling result diagram in one embodiment of the adaptive switching method for intelligent scheduling of an energy system according to the present invention.

[0036] Figure 3This is a schematic diagram of typical monthly system risk results and thresholds in one embodiment of the adaptive switching method for intelligent scheduling of an energy system according to the present invention. Detailed Implementation

[0037] See Figures 1-3 As shown, an adaptive switching method for intelligent scheduling of an energy system includes:

[0038] Initialize sub-models, including artificial neural network agent models and running priority models, as scheduling candidates;

[0039] Based on the forecast data, the net load forecast value is obtained, and the forecast uncertainty index, net load index, and risk index are calculated.

[0040] Calculate the comprehensive risk index by weighted summation of all indicators, and compare it with a preset threshold. If the comprehensive risk index is higher than the threshold, select the high-risk model; otherwise, select the non-high-risk model.

[0041] A scheduling scheme is generated based on the selected model, and the feasibility of the model scheme is verified. If the operational constraints are violated, the scheme is switched to another model scheme and verified again.

[0042] Existing energy system scheduling often relies on a single algorithm. While traditional Gurobi solver-based modeling algorithms can yield relatively good energy system scheduling results, their computational time is long, making it difficult to meet real-time scheduling requirements. Priority-based scheduling algorithms can respond quickly and strictly meet operational constraints, but their global optimization capabilities are limited, easily leading to increased wind and solar curtailment and higher operating costs. Therefore, a single algorithm has significant shortcomings in different operating scenarios, making it difficult to simultaneously achieve real-time performance, economic efficiency, and security.

[0043] Therefore, the technical problem to be solved by this invention is: how to introduce a mechanism in energy system scheduling that can dynamically switch different scheduling algorithms according to the operating scenario, so that the system can achieve fast and economical scheduling under stable operating conditions, and ensure operational safety and feasibility under fluctuating or extreme operating conditions, thereby taking into account real-time performance, economy and safety at the same time.

[0044] like Figure 1 In one embodiment shown, Figure 1 This is an overall flowchart of an adaptive switching method for intelligent energy system scheduling according to the present invention. First, the invention initializes two core scheduling models as candidate schemes: an artificial neural network proxy model that prioritizes economy and speed, and an operation priority model that ensures safety and feasibility. Next, the system calculates the net load forecast based on future wind and solar power output probability prediction data and load prediction data, laying the foundation for subsequent risk assessment.

[0045] After data preparation, a multi-level risk assessment phase begins. The first step is to calculate the prediction uncertainty index. This index quantifies the accuracy of the prediction by dividing the predicted wind and solar power output range for the future target time step by its installed capacity; a higher index value indicates greater uncertainty. Subsequently, this invention assesses the net load ramp-up risk index. Specifically, it first calculates the net load change rate for adjacent future time steps and then compares it to the system's allowable regulation rate, calibrated by the comprehensive regulation capacity of non-renewable energy generator units and energy storage equipment. The closer the ratio is to 1, the greater the system regulation pressure. Simultaneously, to provide early warning of extreme power supply and demand imbalances, separate indicators for power shortage risk and power curtailment risk are calculated: For power shortage risk, the estimated net load is calculated using the lower limit of the wind and solar power output prediction under pessimistic conditions. If this value exceeds a preset proportional threshold for the upper limit of non-renewable energy regulation unit output, a risk is identified. For power curtailment risk, the estimated net load is calculated using the upper limit of the wind and solar power output prediction under optimistic conditions. If this value is lower than a preset proportional threshold for the lower limit of non-renewable energy regulation unit output, a risk is identified. In addition, a net load volatility index, which characterizes the degree of fluctuation, is obtained by calculating and normalizing the square of the difference between net load forecasts at adjacent time points.

[0046] After obtaining all the above indicators, the next step is to make a comprehensive decision. Each indicator value is multiplied by its weight, obtained through offline calibration and optimization, and then summed to obtain a comprehensive risk index. The weights and comparison thresholds here are optimal parameters determined in advance based on historical full-year data, with the goal of minimizing the total annual operating cost, including fossil fuel costs, equipment maintenance costs, and curtailment penalty costs, through enumeration or search methods. Next, this comprehensive risk index is compared with the optimized upper threshold. If the comprehensive risk index is higher than this threshold, it is determined to be a high-risk scenario, and the system will prioritize the execution priority model to generate a scheduling plan; otherwise, the artificial neural network proxy model will be selected.

[0047] Then, after generating candidate scheduling schemes, there is a crucial post-verification step. Especially when selecting an artificial neural network surrogate model, it is essential to rigorously check whether the scheme satisfies all operational constraints, including the upper and lower limits of unit output, energy storage regulation boundaries, and power balance constraints. If a scheme is found to be infeasible, the system will immediately switch to the scheme generated by the priority model to ensure the absolute safety and reliability of the scheduling results.

[0048] Finally, it is important to emphasize that the framework of this invention is highly scalable. Its core lies in the risk index and threshold determination mechanism, and therefore it is not limited to the two models mentioned above. By subdividing the risk range, other algorithms such as robust optimization models and heuristic search models can be incorporated into this dynamic switching system, thus flexibly adapting to more scheduling needs in the future.

[0049] In one embodiment, the specific steps of the present invention are as follows:

[0050] Step 1: Scheduling algorithm preparation, rolling time domain setting, and data preparation

[0051] Prepare multiple scheduling optimization algorithms:

[0052] (1) Artificial neural network proxy model algorithm: It is trained by historical Gurobi optimization results and can output an approximate optimal solution in milliseconds. It can balance economy and computing speed. However, this solution method is prone to scheduling schemes that do not meet physical constraints and may fail in complex and extreme cases.

[0053] (2) Setting the running priority algorithm: The priority logic is set based on the physical constraints of energy equipment. The main advantages are safety and feasibility. The disadvantage is that due to the setting of priorities, the scheduling strategy is relatively fixed and lacks optimization, making it difficult to achieve economic optimization.

[0054] It can be extended to third or more algorithms (such as heuristics and robust optimization) without changing the methodological framework.

[0055] Set the scheduling step size (e.g., 15 minutes). At each time t, obtain the load forecast for the next time t+1 and the probability forecast range of renewable energy (wind power, solar power) output (upper and lower limits of the forecast interval).

[0056] Step 2: Predict Net Load Values

[0057] By combining the wind power, solar power, and load forecast values ​​(load demand - renewable energy generation), the net load of the power system at future times is obtained, which is the value of power surplus or shortage. The net load forecast value of time t to the future time t+1 can be obtained by subtracting the wind and solar power output value of time t to the future time t+1 from the load forecast value of time t to the future time t+1.

[0058] Step 3: Predict the Uncertainty Indicator (UI)

[0059] This step primarily measures the accuracy of wind and solar power forecasts. If complex weather conditions result in a wider forecast range for wind and solar power output, it indicates greater uncertainty, making it less suitable to use a relatively unstable surrogate model. The forecast uncertainty index (UI) can be obtained as follows:

[0060] (1) First, obtain the predicted range of wind and solar power output at the (t+1)th time step in the future, that is, the upper limit value minus the lower limit value of the prediction;

[0061] (2) Then divide the wind and solar power output prediction range by the wind and solar installed capacity in the energy system to obtain the prediction uncertainty index (UI).

[0062] Step 4: Net Load Ramp-Up Risk (RRI)

[0063] This step primarily assesses the rate of change in future net load to measure the degree of net load volatility. If the volatility of net load approaches the system's safe adjustment speed too rapidly, an artificial neural network surrogate model is not suitable. The specific calculations are as follows:

[0064] (1) First, it is necessary to evaluate the net load change rate at time step t+1. This is done by subtracting the net load value predicted at time step t+1 from the net load value predicted at time step t+2, taking the absolute value, and then dividing by the time interval Δt. This gives the net load change rate at time step t+1 in the future.

[0065] (2) By comparing the rate of change of net load at time t+1 in the future with the allowable adjustment rate of the system (which can be calibrated by fossil energy generator sets + energy storage capacity), the system ramp-up risk RRI can be obtained.

[0066] If the RRI value is close to 1, it means that the required regulation speed of the energy system is close to the safe regulation speed that the system can achieve, which poses a certain challenge to the safe operation of the system and makes it difficult to adopt artificial neural network surrogate model methods.

[0067] Step 5: Power Shortage / Abandonment Risk Detection (SPI / CPI)

[0068] This step is used to determine in advance whether there is a risk of power shortage or power curtailment in the next step.

[0069] The method for determining the risk of power shortage (SPI) is as follows: when the forecast results indicate that the output of wind and solar power may be low, take the estimated value of the next net load under the pessimistic scenario (i.e., the net load calculated by removing the lower limit of the wind and solar power output forecast). If this value exceeds a certain percentage threshold of the upper limit of the output of the system's fossil energy regulating units (e.g., 0.8 times the maximum output of the controllable fossil energy regulating units), then it is determined that there is a risk of power shortage, and SPI is recorded as 1; otherwise, it is determined that there is no risk of power shortage, and SPI is recorded as 0.

[0070] The method for determining the curtailment risk (CPI) is as follows: when the forecast results show that the output of wind and solar power may be high, take the estimated value of the next net load under the optimistic scenario (i.e., the net load calculated by taking the upper limit of the wind and solar power output forecast). If the value is lower than a certain percentage threshold of the lower limit of the output of fossil energy regulating units (e.g., 1.2 times the minimum output of fossil energy regulating units), then it is determined that there is a curtailment risk, and the CPI is recorded as 1; otherwise, it is determined that there is no curtailment risk, and the CPI is recorded as 0.

[0071] When SPI=1, it indicates that the system may experience insufficient power; when CPI=1, it indicates that the system may experience wind and solar power curtailment.

[0072] Step 6: Assess Net Load Volatility (VI)

[0073] To assess the volatility (VI) of net load over a future period, this invention establishes a volatility index. This index is calculated by comparing the net load forecasts at two adjacent time points, squared the difference, and normalized using the installed capacity of renewable energy within the system as a benchmark, thus obtaining the relative volatility intensity. Here, the net load forecast refers to the prediction of the net load at a given time point for the next time point; using the installed capacity of renewable energy as the denominator makes the volatility levels comparable between systems of different sizes. A higher volatility index indicates more drastic changes in future net load, resulting in higher scheduling difficulty. Therefore, conservative scheduling methods that prioritize safety should be prioritized when selecting algorithms.

[0074] Step 7: Setting the switching rules for the scheduling optimization algorithm

[0075] Based on the evaluation parameters set in steps three through six, construct the evaluation index S. t To assess the overall operational risk of the energy system, and to set upper and lower thresholds (T). high T low To determine which scheduling algorithm is optimal for time t.

[0076] The specific calculation method is as follows: First, obtain the prediction uncertainty index, net load ramp-up risk index, power shortage risk index, power curtailment risk index, and load volatility index. Then, assign a non-negative weight (w1-w5) to each index, and the sum of all weights is 1. Multiply each index value by its corresponding weight and add them together to obtain a comprehensive risk index. The larger the value, the higher the system operation risk.

[0077] During the scheduling algorithm switching process, the comprehensive risk index S is used. t The system is compared with two pre-set thresholds: if the overall risk is greater than the upper threshold, it indicates that the system operation risk is high, and the scheduling method based on setting scheduling priorities is selected to ensure the safety and feasibility of the energy system; if the overall risk index is less than the upper threshold, the scheduling method based on the proxy model is selected to pursue economy and computing speed.

[0078] Step 8: Generate candidate solutions and verify their feasibility.

[0079] At time t, by following steps one through seven and combining wind and solar power output forecasting and load forecasting data, two scheduling output schemes for the scheduling model can be obtained: u t NN (The scheduling strategy at time t given by the artificial neural network surrogate model) and u t RL(Set the scheduling strategy at time t given by the priority model). If the artificial neural network proxy model method is recommended after calculations in steps two through seven, it is also necessary to determine whether it meets various constraints. If the scheduling scheme given by the artificial neural network proxy model violates various restrictions on system operation (such as the scheduling scheme exceeding the upper and lower limits of unit output, energy storage regulation, etc.), then the scheduling scheme result with the set priority will be used immediately at time t.

[0080] If the artificial neural network scheduling method is selected, it is necessary to determine whether the artificial neural network scheduling scheme meets the system's operational constraints. If the priority scheduling scheme should be selected at time t, no further judgment is needed. Generally, under special operating conditions, the artificial neural network scheduling scheme may fail to meet the constraints. For example, if the power generation of renewable energy suddenly decreases, energy storage may be needed to quickly generate electricity to make up for it. Such a large-scale adjustment may cause the optimization result generated by the artificial neural network to exceed the maximum amount of electricity that the energy storage can release.

[0081] Step 9: Offline calibration of weights and thresholds

[0082] To assign weights (w1-w5) to the risk assessment parameters and the algorithm switching threshold T high To find the most suitable settings that enable stable system operation, minimize power wastage, and reduce costs, this invention simulates the scheduling process based on historical data, combines steps one through eight, accumulates the operating costs at each time step, uses minimizing the operating cost as the objective function, continuously changes the parameter values, and finally selects the parameter values ​​from the scheme with the lowest operating cost as the setting results for weights and thresholds.

[0083] When constraining the parameters, firstly, it is stipulated that each weight (w1-w5) is a non-negative number, and the sum of all weights equals one; secondly, it is stipulated that the algorithm switches to the upper threshold (T). high (The value is less than or equal to one.) By enumerating or searching, each possible combination of weights and thresholds is simulated to obtain a set of candidate solutions N, from which the parameter setting scheme with the lowest total cost is selected.

[0084] Inputting the weight combination and threshold combination scheme N* for each group, and based on the historical annual data of the energy system (including renewable energy output data and load data), the operating cost for each time step can be calculated through simulation using steps one through eight mentioned above. The operating cost consists of three parts: the first part is the fossil fuel usage cost, i.e., the cost corresponding to the fuel consumed by fossil fuel units at that time step; the second part is the equipment operation and maintenance cost, which is the cost obtained by multiplying the installed capacity of the equipment by the unit operation and maintenance cost, and can be calculated based on the total system capacity or separately for each unit; the third part is the curtailment penalty cost, i.e., the amount of electricity curtailed at that time step multiplied by the unit curtailment penalty cost. Adding these three parts together yields the total operating cost for that time step.

[0085] After obtaining the operating cost for each time step throughout the year, the operating costs for all time steps are summed to obtain the total annual operating cost J(N*) for the weight and threshold combination scheme N*. For each combination of weight and threshold parameters, the corresponding total annual operating cost J can be calculated using the above method. Finally, among all candidate schemes, the set of parameters with the minimum total annual operating cost is selected as the optimal setting values ​​w* and T for the risk assessment weights and algorithm switching threshold. high *

[0086] Furthermore, the scheduling framework of this invention is not limited to neural networks and preset priority algorithms, but can be extended to various other scheduling algorithms. First, a comprehensive risk index is calculated using historical data. Then, one or more thresholds are set to map different intervals to different scheduling algorithms. For example, when the risk index is low, an economical algorithm such as a surrogate model can be used; when the risk index exceeds the set threshold, a more robust algorithm such as rule priority, robust optimization, or heuristic search is switched to. If support for three or more algorithms is required, the risk intervals can be further subdivided, with each interval corresponding to a different algorithm. The threshold values ​​are still determined through historical simulation and minimizing operating costs, so effective switching can be achieved within this unified framework regardless of the algorithm type extended to.

[0087] like Figure 2 and Figure 3 In one embodiment shown, Figure 2 This is a typical monthly system scheduling result diagram in one embodiment of the adaptive switching method for intelligent energy system scheduling according to the present invention. Figure 3 This is a schematic diagram illustrating typical monthly system risk results and thresholds in one embodiment of the adaptive switching method for intelligent energy system scheduling according to the present invention. Figure 2The pink portion represents the artificial neural network scheduling method used in this invention, while the white portion represents the conventional scheduling method used instead. Simulations show that when the system operates under normal conditions with stable wind and solar power output, slow load changes, and small prediction errors, the artificial neural network proxy scheduling method is more suitable. In this case, the proxy model can provide an approximate optimal solution in milliseconds, ensuring real-time operation while reducing fuel consumption and computational costs. However, when the system faces drastic fluctuations in wind and solar power output, rapid load increases or decreases, or significantly increased prediction errors, switching to a fixed scheduling logic is more suitable. Although this type of logic is less economical, it strictly adheres to power balance and equipment constraints, ensuring that energy storage and diesel engines can step in promptly, thereby guaranteeing the safety and feasibility of the system during critical periods.

[0088] The advantages of this invention are:

[0089] 1. Instead of choosing a fixed scheduling method, the system dynamically determines whether to use a neural network agent or a priority algorithm based on the volatility and uncertainty of wind and solar forecasts. This allows the system to pursue economy in stable scenarios and ensure safety in volatile or extreme scenarios, fully leveraging the complementary advantages of different algorithms.

[0090] 2. The scheduling framework proposed in this invention is not limited to neural networks and preset priority algorithms, but can be extended to various scheduling methods including robust optimization, heuristic search, and simulation-driven control. Because the overall structure and specific algorithm implementation are decoupled, the framework has high versatility and evolvability. Extension to other algorithms can still be achieved by switching algorithms using a "risk index + threshold determination" approach.

Claims

1. An adaptive switching method for intelligent scheduling of an energy system, characterized in that, include: Initialize sub-models, including artificial neural network agent models and running priority models, as scheduling candidates; Based on the forecast data, the net load forecast value is obtained, and the net load index, risk index, and forecast uncertainty index are calculated to quantify the accuracy of wind and solar forecasts. All indicators are weighted and summed to calculate a comprehensive risk index, which is then compared with a preset threshold. If the comprehensive risk index is higher than the threshold, the priority model is selected; otherwise, the artificial neural network proxy model is selected. A scheduling scheme is generated based on the selected model, and the feasibility of the model scheme is verified. If the operational constraints are violated, the scheme is switched to another model scheme and verified again.

2. The adaptive switching method for intelligent scheduling of an energy system according to claim 1, characterized in that, The net load indicators include net load ramp-up risk indicators and net load volatility indicators; the risk indicators include power shortage risk indicators and power curtailment risk indicators; the forecast data includes wind and solar power output probability forecast data and load forecast data for future times.

3. The adaptive switching method for intelligent scheduling of an energy system according to claim 2, characterized in that, The calculation of the prediction uncertainty index specifically includes: obtaining the wind and solar power output prediction range for the future target time step, which is obtained by subtracting the lower limit value from the upper limit value of the wind and solar power output prediction; dividing the wind and solar power output prediction range by the wind and solar installed capacity in the energy system to obtain the normalized prediction uncertainty index; the larger the prediction uncertainty index value, the higher the prediction uncertainty, and the less suitable it is to use an artificial neural network surrogate model.

4. The adaptive switching method for intelligent scheduling of an energy system according to claim 2 or 3, characterized in that, The calculation of the net load ramp-up risk index specifically includes: assessing the net load change rate at the future target time step, which is obtained by subtracting the net load value at the future first time step from the net load value predicted at the current moment, taking the absolute value, and then dividing by the time interval; then comparing the net load change rate with the system's allowable adjustment rate, which is calibrated by the comprehensive adjustment capability of non-renewable energy generator sets and energy storage devices; the closer the net load ramp-up risk index value is to 1, the closer the net load fluctuation rate is to the system's safety limit, and the less suitable it is to use an artificial neural network proxy model.

5. The adaptive switching method for intelligent scheduling of an energy system according to claim 4, characterized in that, The power shortage risk indicators and power curtailment risk indicators specifically include: For the power shortage risk indicator, when the predicted wind and solar power output is low, the estimated net load calculated using the lower limit of the predicted wind and solar power output under pessimistic conditions is used. If this value exceeds the preset proportional threshold of the upper limit of the output of non-renewable energy regulating units, then a power shortage risk is determined, and the indicator is marked as 1; otherwise, it is marked as 0. For the power curtailment risk indicator, when the predicted wind and solar power output is high, the estimated net load calculated using the upper limit of the predicted wind and solar power output under optimistic conditions is used. If this value is lower than the preset proportional threshold of the lower limit of the output of non-renewable energy regulating units, then a power curtailment risk is determined, and the indicator is marked as 1; otherwise, it is marked as 0.

6. The adaptive switching method for intelligent scheduling of an energy system according to claim 5, characterized in that, The net load volatility index specifically includes: calculating the square of the difference between the net load forecast values ​​at adjacent time points, and normalizing it based on the installed capacity of renewable energy. The larger the volatility index value, the more drastic the future net load changes, the higher the scheduling difficulty, and the more suitable it is to prioritize the operation priority model.

7. The adaptive switching method for intelligent scheduling of an energy system according to claim 6, characterized in that, When calculating the comprehensive risk index using weighted summation, the weight values ​​are determined through offline calibration and optimization; the preset threshold includes an upper threshold, which is optimized through historical data simulation with the goal of minimizing the annual operating cost; The comprehensive risk index is the sum of the values ​​of each indicator multiplied by their corresponding weights. The comparison between the index and the threshold distinguishes between high-risk and low-risk scenarios.

8. The adaptive switching method for intelligent scheduling of an energy system according to claim 7, characterized in that, The feasibility verification specifically includes: when an artificial neural network proxy model is selected, checking whether the scheduling scheme it generates meets the system operation constraints, including the upper and lower limits of unit output, the energy storage regulation boundary, and the power balance constraints; if any constraint is violated, the scheduling scheme generated by the operation priority model is switched to.

9. The adaptive switching method for intelligent scheduling of an energy system according to claim 7, characterized in that, The weight values ​​are optimized through offline calibration, which includes: simulating the scheduling process based on historical full-year data and calculating the operating cost at each time step. The operating cost includes fossil fuel costs, equipment maintenance costs, and power curtailment penalty costs. With the goal of minimizing the total annual operating cost, the combination of weight and threshold parameters is optimized through enumeration or search, and the parameter with the lowest total cost is selected as the set value.

10. The adaptive switching method for intelligent scheduling of an energy system according to claim 1, characterized in that, The specific method for switching to other models includes subdividing several models by risk index range. When the comprehensive risk index is in different ranges, it corresponds to different scheduling models, including robust optimization models and heuristic search models.

Images