Multi-time scale scene analysis-based day-ahead and intra-day optimal scheduling method for micro-grid

Abstract

The invention relates to a multi-time scale scene analysis-based micro-grid day-ahead and intra-day optimal scheduling method, which belongs to the field of micro-grid scheduling, and is characterized in that a variational mode decomposition (VMD)-long short-term memory network (LSTM) multi-scale prediction framework is constructed, ultra-short-term precision is improved through variable mode decomposition and frequency division prediction, a Canopy-spectral clustering-K-means hybrid algorithm is designed, and the optimal scheduling of a micro-grid is realized. A typical scene is generated based on Latin hypercube sampling (LHS), the scene coverage capability is enhanced, a day-ahead and intra-day two-stage optimization model is finally constructed, a high-dimensional problem is rapidly solved by adopting a mixed integer programming algorithm, and theoretical support is provided for high-proportion renewable energy consumption and micro-grid refined scheduling.

Description

TECHNICAL FIELD

[0001] The application relates to a micro-grid day-ahead and day-ahead optimization scheduling method based on multi-time scale scene analysis, and belongs to the micro-grid scheduling field. BACKGROUND

[0002] With the continuous growth of the penetration rate of renewable energy such as wind power and photovoltaic power in a micro-grid, the output of the renewable energy is affected by meteorological and space-time factors, presents strong fluctuation and anti-peaking characteristics, and leads to the aggravation of the source-load power imbalance risk, thereby posing a severe challenge to the scheduling economy and reliability. The micro-grid can effectively smooth wind and light fluctuations by coordinating distributed power sources, energy storage and controllable units. However, the existing research still has the following problems: 1. Traditional prediction methods (such as EMD and ARIMA) cannot separate the multi-scale characteristics of wind and light output, thereby limiting the accuracy of ultra-short-term prediction; 2. Scene generation is mainly dependent on single clustering algorithms such as K-means, and the distance between classes is low and the scene coverage is insufficient; and 3. Optimization models often ignore the coupling mechanism of two-stage constraints, and heuristic algorithms are prone to local optimum in high-dimensional problems. SUMMARY

[0003] According to the problems described in the background, the application aims to provide a micro-grid day-ahead and day-ahead optimization scheduling method based on multi-time scale scene analysis. A VMD-LSTM multi-scale prediction framework is constructed to improve the accuracy of ultra-short-term prediction through variable mode decomposition and frequency prediction. A Canopy-spectral clustering-K-means hybrid algorithm is designed to generate typical scenes based on Latin hypercube sampling (LHS) to enhance the scene coverage. Finally, a day-ahead and day-ahead two-level optimization model is constructed, and a mixed integer programming algorithm is used to realize fast solving of high-dimensional problems, thereby providing theoretical support for high proportion of renewable energy consumption and fine scheduling of micro-grids, and solving the above problems.

[0004] To achieve the above purpose, the application provides the following technical scheme: a micro-grid day-ahead and day-ahead optimization scheduling method based on multi-time scale scene analysis is provided, which comprises the following steps:

[0005] (1) A VMD-LSTM multi-scale prediction framework is constructed to improve the accuracy of ultra-short-term prediction through variable mode decomposition and frequency prediction;

[0006] (2) A Canopy-spectral clustering-K-means hybrid algorithm is designed to generate typical scenes based on LHS to enhance the scene coverage;

[0007] (3) A day-ahead and day-ahead two-level optimization model is constructed, and a mixed integer programming algorithm is used to realize fast solving of high-dimensional problems.

[0008] Preferably, the step (1) comprises the following steps:

[0009] (1.1) Extracting multi-scale features of wind-solar hybrid power processing by VMD;

[0010] (1.2) Making scale prediction by LSTM, the prediction process including three stages: forgetting stage, memory selection stage and output stage;

[0011] (1.3) Building VMD-LSTM algorithm process, forming a multi-scale prediction framework.

[0012] Preferably, the VMD used in the step (1.1) adaptively decomposes the original output sequence f(t) into K intrinsic mode functions (IMF) by constructing a constrained variational problem, and the mathematical expression is:

[0013]

[0014] The constraint condition is:

[0015]

[0016] In the formula, u k (t) is the kth IMF component, representing different frequency domain features; ω k is the center frequency of the corresponding component; f(t) is the original wind-solar output sequence; denotes the time derivative, δ(t) is the Dirac function, and j is the imaginary unit;

[0017] VMD can effectively eliminate mode aliasing and extract multi-scale features of wind-solar output by iterative solution through alternating direction multiplier method (ADMM).

[0018] Preferably, the algorithm process in the step (1.3) comprises:

[0019] (1.3.1) Using VMD algorithm to process wind-solar output data under time series decomposition, obtaining modal components IMFn with different fluctuation characteristics;

[0020] (1.3.2) Using LSTM neural network to predict the future trend of each modal component;

[0021] (1.3.3) Adding each modal prediction component to calculate the final wind-solar prediction data.

[0022] Preferably, the step (2) comprises the following steps:

[0023] (2.1) Using LHS to generate scenarios;

[0024] (2.2) Improved Canopy-spectral clustering-K-means hybrid algorithm, reduce the scene dimension and extract typical features.

[0025] Preferably, LHS is used in step (2.1) to generate a scene, 1000 random scenes are generated in the wind and light output probability distribution space, and LHS ensures uniform coverage of each dimension through stratified sampling, which is mathematically expressed as:

[0026] The wind and light output probability distribution function F(x) is stratified, and N=1000 equal probability intervals are divided;

[0027] A sample point x is randomly extracted in each interval i , satisfying:

[0028]

[0029] In the formula, F -1 is the inverse function of the cumulative distribution function, and ξ i is a uniformly distributed random number;

[0030] Among them, the wind power fluctuation range under the day-ahead scenario is ±30%, and the photovoltaic is ±25%, covering typical weather modes (sunny, cloudy, cloudy); The intra-day scenario takes 15 minutes as a step, simulates the ultra-short-term prediction deviation, and reflects the rapid fluctuation characteristics of wind and light.

[0031] Preferably, the process of improving the Canopy-spectral clustering-K-means hybrid algorithm in step (2.2) includes the following steps:

[0032] (2.2.1) Determine the optimal number of scenes K through Canopy rough clustering;

[0033] (2.2.2) Construct a similarity matrix using spectral clustering and reduce dimension to generate a low-dimensional space matrix Y;

[0034] (2.2.3) Adopt K-means to fine cluster the row vectors of Y matrix, fuse the advantages of three kinds of algorithms, and output K scenes representing typical fluctuation characteristics.

[0035] Preferably, step (3) includes the following steps:

[0036] (3.1) Construct a day-ahead-intra-day two-level optimization framework;

[0037] (3.2) Multi-objective model and constraint condition establishment.

[0038] Preferably, the optimization framework process in step (3.1) includes:

[0039] (3.1.1) Based on the Variational Mode Decomposition (VMD) to generate wind and solar power output short-term prediction curve, through Latin hypercube sampling to construct random scene set, and using Canopy-spectral clustering-K-means hybrid algorithm to reduce scene, and extract typical output mode;

[0040] (3.1.2) Establishing a 24-hour day-ahead scheduling model, taking typical scenes as input, optimizing micro gas turbine, energy storage and grid interactive power plan with 1-hour time resolution, and solving the comprehensive optimal solution of economy and environment throughout the day;

[0041] (3.1.3) Based on the 15-minute intra-day short-term prediction data, a rolling optimization model is constructed, the scheduling instruction is updated every 1 hour, the power difference caused by the deviation of wind and solar prediction is dynamically corrected, the minute-level real-time response is realized through mixed integer programming (MILP), and the coordination optimization of day-ahead plan and real-time operation is ensured.

[0042] Preferably, the multi-objective model in step (3.2) is divided into a day-ahead stage multi-objective model and an intra-day stage dynamic correction model, and the constraint conditions include power balance constraint and energy storage dynamic model;

[0043] The economic target of the day-ahead stage multi-objective model is:

[0044]

[0045] In the formula, T=24 is the total number of scheduling periods; C WT-PV,t is the comprehensive cost of wind and solar power generation in t period (including energy penalty and operation and maintenance cost, unit: yuan); C MT,t is the micro gas turbine fuel and maintenance cost; C ESS,t is the energy storage charging and discharging loss cost; C Grid,t is the grid interaction cost (positive for power purchase, negative for power sale);

[0046] The environmental benefit target of the day-ahead stage multi-objective model is:

[0047]

[0048] In the formula, j=3 is the type of pollutants (CO2, SO2, NOx); λ j is the treatment cost of the jth type of pollutants (unit: yuan / kg); E j,t is the micro gas turbine emission amount (kg) in t period, which is calculated by output P MT,t and emission coefficient;

[0049] The intra-day stage dynamic correction model is:

[0050] minF3=αΔC+(1-α)(|ΔP Grid |+|ΔP MT|+|ΔP ESS

[0051] Wherein, alpha=0.7 is the cost deviation weight; Delta C is the deviation of real-time operation cost and day-ahead plan (yuan); Delta P Grid , Delta P MT , Delta P ESS Respectively, the grid interaction power, the micro gas turbine output and the adjustment amount of energy storage (kW);

[0052] The power balance constraint is:

[0053] P WT,t +P PV,t +P MT,t +P Grid,t +P ESS,t =P Load,t

[0054] Wherein, P WT,t , P PV,t Are the wind power and photovoltaic output (kW) in t period; P MT,t Is the micro gas turbine output (kW);P Grid,t Is the grid interaction power (positive for power purchase, negative for power sale);P ESS,t Is the energy storage charging and discharging power (positive for discharging, negative for charging);P Load,t Is the load demand (kW);

[0055] The energy storage dynamic model is:

[0056]

[0057] Wherein, SOC t Is the energy storage state of charge in t period;Eta ch , Eta dis Are the charging and discharging efficiency;E ESS Is the rated capacity of energy storage;Delta t is the time interval.

[0058] The beneficial effects of the application are:

[0059] 1、Construct VMD-LSTM multi-scale prediction framework, improve ultra-short-term accuracy through modal decomposition and frequency division prediction, wind power intra-day prediction MAE reduces to 0.60, which is 44% higher than day-ahead prediction.

[0060] 2、Design Canopy-spectrum clustering-K-means hybrid algorithm to generate typical scenarios, the distance between classes is improved by 43% compared with traditional K-means, which significantly enhances the scenario coverage ability.

[0061] ​3、Establish day-to-day optimization model, day-to-day stage to generate benchmark plan with economic-environmental cost collaborative optimization as target, and compare NSGA-II and improved PSO algorithm performance. The example shows that, the total running cost of MILP algorithm is reduced by 8.3% and 12.2% respectively compared with NSGA-II and improved PSO, and the solving efficiency is improved by 8 times; the controllable equipment adjustment amount is reduced by 47% after day-to-day optimization, and the abandoned wind and light rate is stable at 0.9%, which verifies the effectiveness of the proposed strategy in improving economy, reducing carbon emission and enhancing system robustness. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 The step schematic diagram of the application is shown in the figure;

[0063] Figure 2 The day-to-day photovoltaic power center frequency iteration curve and fitting error graph when K is different are shown in the figure;

[0064] Figure 3 The wind and light prediction flow chart based on variational mode decomposition is shown in the figure;

[0065] Figure 4 The wind and light scene analysis flow chart based on improved clustering reduction is shown in the figure;

[0066] Figure 5 The clustering result probability distribution comparison graph is shown in the figure;

[0067] Figure 6 The inter-class distance comparison graph is shown in the figure;

[0068] Figure 7 The microgrid day-to-day-intra-day optimization scheduling flow chart is shown in the figure;

[0069] Figure 8 The microgrid load prediction graph is shown in the figure;

[0070] Figure 9 The day-to-day scheduling result graph based on classical mixed integer programming solution is shown in the figure;

[0071] Figure 10 The day-to-day scheduling result solved by NSGA-II algorithm is shown in the figure;

[0072] Figure 11 The day-to-day scheduling result solved by improved PSO algorithm is shown in the figure;

[0073] Figure 12 The day-to-day scheduling result based on classical mixed integer programming is shown in the figure;

[0074] Figure 13 The day-to-day scheduling result solved by NSGA-II algorithm is shown in the figure;

[0075] Figure 14 the intraday scheduling result solved by the improved PSO algorithm; DETAILED DESCRIPTION

[0076] The embodiments of the present application will be further described below with reference to the drawings:

[0077] Embodiment 1

[0078] As shown in the figure, the present application provides a micro-grid day-ahead and intraday optimization scheduling method based on multi-time scale scene analysis, including the following steps: Figure 1

[0079] (1) Construct a VMD-LSTM multi-scale prediction framework to improve ultra-short-term accuracy through variable mode decomposition and frequency division prediction;

[0080] (1.1) Extract multi-scale features of wind and light processing through VMD;

[0081] The wind power and photovoltaic output are affected by weather conditions and spatial and temporal distribution, showing significant non-stationary and multi-scale fluctuation characteristics. The traditional signal decomposition method such as empirical mode decomposition (EMD) is prone to modal aliasing, which limits the prediction accuracy. Therefore, the variational mode decomposition (VMD) is used to construct a constrained variational problem to adaptively decompose the original output sequence f(t) into K intrinsic mode functions (IMF). The mathematical expression is:

[0082]

[0083] The constraint condition is:

[0084]

[0085] In the formula, u k (k) is the kth IMF component, representing different frequency domain characteristics; ω k is the center frequency of the corresponding component; f(t) is the original wind and light output sequence; denotes the time derivative, δ(t) is the Dirac function, and j is the imaginary unit;

[0086] VMD can effectively eliminate modal aliasing and extract multi-scale features of wind and light output by iterative solution through the alternating direction multiplier method (ADMM).

[0087] Taking the day-ahead photovoltaic output as an example, the optimal modal number K = 4 is determined by the center frequency convergence curve, and the decomposition result is as shown in the figure. Figure 2

[0088] (1.2) Perform scale prediction through LSTM, which includes three stages: forgetting stage, memory selection stage and output stage;​​

[0089] The forgetting phase is mainly to selectively eliminate the information passed from the previous node. By reading the output state of the previous node and the output state of the current node, the activation function determines the proportion of the cell state information forgotten at the previous moment;

[0090] The selective memory phase is to selectively filter the input information and select important information to input to the current cell;

[0091] The selective memory phase mainly includes two steps: the first step is to determine which information needs to be updated by the activation function of the input layer, and the layer generates a candidate update content; The second step is to update the cell state, and the output state of the forgetting phase and the candidate update cell state are used to obtain the current cell state;

[0092] The output phase will use the activation function to determine the proportion of the output cell state, and then multiply the cell state processed by the layer with the output cell state to obtain the final output result.

[0093] (1.3) Construct the VMD-LSTM algorithm flow to form a multi-scale prediction framework.

[0094] Considering the uncertainty characteristics of wind and light processing, first, the variational mode is used to decompose the component curve with typical characteristics, and then the indirect prediction of wind and light output data is carried out by predicting each component and integrating the method, so as to improve the accuracy of subsequent neural network prediction.

[0095] The flow chart is shown in Figure 3 ;

[0096] The algorithm flow includes:

[0097] (1.3.1) Use VMD algorithm to process wind and light output data under time series decomposition to obtain modal components IMFn with different fluctuation characteristics;

[0098] (1.3.2) Use LSTM neural network to predict the future trend of each modal component;

[0099] (1.3.3) Add the predicted components of each modal to calculate the final wind and light prediction data.

[0100] The LSTM model is trained for the decomposed IMF component, and the day-ahead prediction time scale is 1 hour, and the intra-day prediction scale is 15 minutes. The experiment shows that the ultra-short-term prediction accuracy is significantly improved, and the average absolute error (MAE) of wind power intra-day is 0.60, which is reduced by 44% compared with day-ahead prediction. The experimental results are shown in Table 1:

[0101] Table 1 Comparison of wind and light day-ahead and intra-day output prediction errors

[0102]

[0103] (2) Design Canopy-spectral clustering-K-means hybrid algorithm, generate typical scenarios based on LHS, and enhance the coverage ability of scenarios;

[0104] (2.1) Use LHS to generate scenarios;

[0105] The output of wind power and photovoltaic power is affected by the randomness of weather and equipment, and its uncertainty needs to be characterized by probability method, so LHS is used to generate scenarios, 1000 random scenarios are generated in the probability distribution space of wind and light output, and LHS ensures that samples are evenly covered in each dimension through stratified sampling, which is mathematically expressed as:

[0106] Layer the wind and light output probability distribution function F(x) to divide N=1000 equal probability intervals;

[0107] Randomly extract a sample point x i in each interval, which satisfies:

[0108]

[0109] In the formula, F -1 is the inverse function of the cumulative distribution function, and ξ i is a uniformly distributed random number;

[0110] Among them, the fluctuation range of wind power output under day-ahead scenario is ±30%, and that of photovoltaic power is ±25%, which covers typical weather modes (sunny, cloudy, cloudy); The intra-day scenario takes 15 minutes as a step to simulate the super short-term prediction deviation and reflect the rapid fluctuation characteristics of wind and light.

[0111] (2.2) Improve the Canopy-spectral clustering-K-means hybrid algorithm to reduce the dimension of scenarios and extract typical features.

[0112] The process of improving the Canopy-spectral clustering-K-means hybrid algorithm is shown in Figure 4 , and includes the following steps:

[0113] (2.2.1) Determine the optimal number of scenarios K through Canopy rough clustering;

[0114] Set the distance threshold T1=1.5σ and T2=0.5σ (σ is the standard deviation of the sample), and preliminarily divide the scenario set S:

[0115] First, randomly select a scenario s i as the Canopy center;

[0116] Then, the distance d(s i ,s j) <T1的场景纳入同一Canopy;

[0117] Finally, remove the distance d(s i ,s j ) <T2的场景,避免重复聚类。

[0118] According to the above process, the Canopy algorithm is run to perform coarse clustering on the sampled samples and determine the appropriate K value for scene clustering. Table 2 shows the calculation results.

[0119] Table 2 Canopy clustering calculation results for scenery scenes

[0120]

[0121] Through coarse clustering, the number of wind power scenarios K is preliminarily determined. wind =6, Photovoltaic K pv =4.

[0122] (2.2.2) Use spectral clustering to construct a similarity matrix and reduce the dimension to generate a low-dimensional space matrix Y;

[0123] Construct a similarity matrix W, whose elements W ij Represents scene s i With s j Similarities:

[0124]

[0125] Where σ is the Gaussian kernel width parameter, which controls the local neighborhood range. Calculate the Laplace matrix L = DW (D is the degree matrix, D ii =∑ j W ij ), extract the first K feature vectors to form a low-dimensional embedding space Y∈R N×K .

[0126] (2.2.3) K-means is used to precisely cluster the row vectors of the Y matrix, integrating the advantages of the three algorithms to output K scenarios that represent typical fluctuation characteristics.

[0127] Perform K-means clustering on the low-dimensional space Y, and the objective function is:

[0128]

[0129] Where C k is the kth cluster, μ k is the cluster center.

[0130] like Figure 4As shown, the wind and light scene analysis model proposed in the application generates a large number of random scenes by Latin hypercube sampling (LHS), and adopts a Canopy-spectral clustering-K-means hybrid algorithm to realize scene reduction.

[0131] First, the optimal scene number K is determined by Canopy coarse clustering, a similarity matrix is constructed by spectral clustering and a low-dimensional space matrix Y is generated by dimension reduction, and finally K-means is used for fine clustering of the row vectors of the Y matrix, the advantages of the three algorithms are fused, and K scenes representing typical fluctuation characteristics are output, providing high-coverage scene input for microgrid optimization scheduling.

[0132] To verify the advantages of the Canopy-spectral clustering-Kmeans clustering scene reduction method proposed in this paper over the traditional K-means algorithm, the following compares the results of the two methods by taking wind power day-ahead scene clustering as an example.

[0133] Table 3 Comparison of DB index of typical scene clustering inter-distance

[0134]

[0135] According to the above table and Figure 5 and Figure 6 As shown, the Canopy-spectral clustering-K-means clustering method proposed in the application can not only objectively obtain the optimal clustering number K when reducing wind and light prediction scenes, but also obtain more concentrated and obvious clustering results. At the same time, it can be seen from the calculation results in the above table that the inter-class distance of the results obtained by the clustering method proposed in this paper is larger, and the difference between classes is more obvious, that is, the clustering effect is better, which verifies the effectiveness of the method proposed in the application.

[0136] (3) Construct a day-ahead and day-ahead two-level optimization model, and use a mixed integer programming algorithm to realize fast solution of high-dimensional problems.

[0137] (3.1) Construct a day-ahead-day-ahead two-level optimization framework;

[0138] To cope with the multi-time scale uncertainty of wind and light output, the application constructs a day-ahead-day-ahead two-level optimization framework. The day-ahead stage takes 24 hours as a cycle, and generates a benchmark scheduling plan with the optimal economic cost and environmental benefit as the target; the day-ahead stage takes 15 minutes as a time window and rolls to correct, and minimizes the real-time operation cost and equipment adjustment amount. The flow chart is as Figure 7 shown.

[0139] The optimization framework process includes:

[0140] (3.1.1) Based on the Variational Mode Decomposition (VMD) to generate wind and solar power output short-term prediction curve, through Latin hypercube sampling to construct random scene set, and using Canopy-spectral clustering-K-means hybrid algorithm to reduce scene, extract typical output mode;

[0141] (3.1.2) Establish a 24-hour day-ahead scheduling model, take typical scene as input, optimize micro gas turbine, energy storage and grid interactive power plan with 1 hour as time resolution, solve the comprehensive optimal solution of economy and environment all day;

[0142] (3.1.3) Based on the 15-minute intra-day short-term prediction data, a rolling optimization model is established, the scheduling instruction is updated every 1 hour, the power difference caused by the deviation of wind and solar prediction is dynamically corrected, the minute-level real-time response is realized through mixed integer programming (MILP), and the coordination optimization of day-ahead plan and real-time operation is ensured.

[0143] (3.2) Multi-objective model and constraint condition establishment.

[0144] The multi-objective model is divided into day-ahead stage multi-objective model and intra-day stage dynamic correction model,

[0145] The economic target of the day-ahead stage multi-objective model is:

[0146]

[0147] In the formula, T=24 is the total number of scheduling periods; C WT-PV,t is the comprehensive cost of wind and solar power generation in t period (including energy penalty and operation and maintenance cost, unit: yuan); C MT,t is the micro gas turbine fuel and maintenance cost; C ESS,t is the energy storage charging and discharging loss cost; C Grid,t is the grid interaction cost (positive for power purchase, negative for power sale);

[0148] The environmental benefit target of the day-ahead stage multi-objective model is:

[0149]

[0150] In the formula, j=3 is the type of pollutants (CO2, SO2, NOx); λ j is the treatment cost of the jth type of pollutants (unit: yuan / kg); E j,t is the emission amount of micro gas turbine in t period (kg), which is calculated by output P MT,t and emission coefficient;

[0151] The intra-day stage dynamic correction model is:

[0152] minF3=αΔC+(1-α)(|ΔP Grid |+|ΔPMT |+|ΔP ESS |

[0153] where, α=0.7 is the cost deviation weight; ΔC is the deviation of real-time operation cost and day-ahead plan (yuan); ΔP Grid , ΔP MT , ΔP ESS are the adjustment amount of grid interaction power, micro gas turbine output and energy storage (kW) respectively;

[0154] The constraint conditions include power balance constraint and energy storage dynamic model.

[0155] The power balance constraint is:

[0156] P WT,t +P PV,t +P MT,t +P Grid,t +P ESS,t =P Load,t

[0157] where, P WT,t , P PV,t are the wind power and photovoltaic output in t period (kW); P MT,t is the micro gas turbine output (kW); P Grid,t is the grid interaction power (positive for power purchase, negative for power sale); P ESS,t is the energy storage charging and discharging power (positive for discharging, negative for charging); P Load,t is the load demand (kW).

[0158] The energy storage dynamic model is:

[0159]

[0160] where, SOC t is the energy storage state of charge in t period; η ch , η dis are the charging and discharging efficiency; E ESS is the energy storage rated capacity; and Δt is the time interval.

[0161] According to the above content, the following example analysis is carried out:

[0162] The micro grid provided in the application includes 800kW wind power, 500kW photovoltaic, 600kW micro gas turbine and 500kW energy storage system, and the grid interaction power limit is ±400kW. The wind and light output data is based on the VMD-LSTM prediction model, and the load curve presents significant peak-valley characteristics (the average value in the valley period is less than 200kW, and the peak value in the peak period is 1.2MW). The pollutant control cost and time-of-use electricity price are shown in Tables 4-7, and the algorithm solution adopts mixed integer programming, NSGA-II and improved PSO.

[0163] Table 4 Wind power, photovoltaic unit capacity

[0164]

[0165] Table 5 Micro power output limit

[0166]

[0167] Table 6 Gas unit pollution gas treatment cost and emission coefficient

[0168]

[0169] Table 7 Time-of-use electricity price

[0170]

[0171] In day-ahead optimal dispatch, the wind and light day-ahead output uses the typical value of the wind and light output prediction scene above, and the micro-grid load uses the predicted value, Figure 8 The load prediction value is given.

[0172] As Figure 8 shown, the micro-grid load curve presents significant peak-valley characteristics: the valley period (0:00-5:00) load average is lower than 200kW, the peak period (14:00-19:00) peak value reaches 1.2MW, and the flat period (8:00-10:00, 19:00-24:00) load demand presents multi-stage fluctuation characteristics. Based on this, mixed integer programming (MILP) and NSGA-II, improved PSO algorithm are used for optimization, and an economic-environmental comprehensive cost objective function is constructed, and the results are shown in Figure 9 、 10 , 11. The specific strategies are as follows:

[0173] 1. Deep valley period (11:00-14:00): wind power is fully supplied (650kW), energy storage is charged at 200kW, micro gas turbine is stopped, the remaining electricity 450kW is sold to the grid at 0.18 yuan / kWh, and SOC is improved to 90%.

[0174] 2. Valley period (0:00-5:00): energy storage charging power is 180kW, micro gas turbine runs at the minimum output of 30kW, grid interaction is relaxed to ±400kW, and fuel cost is reduced by 12.7%.

[0175] 3. Peak period (17:00-19:00): micro gas turbine is short-term over-provisioned to 800kW, energy storage is discharged at 250kW, grid power purchase is limited to 300kW, and total cost accounts for 72%.

[0176] 4. Peak period (14:00-19:00): Micro-turbine stabilizes at 600 kW, energy storage discharges in steps (200 kW→150 kW), demand response reduces 80 kW load, and grid fluctuation rate drops to ±15%.

[0177] 5. Flat period (8:00-10:00, etc.): Energy storage dynamically adjusts ±50 kW threshold, micro-turbine operates at 200 kW economic output, and curtailment rate of wind and solar power drops to 0.9%.

[0178] To respond to real-time fluctuations in wind and solar output and load, the intraday optimization uses 15-minute time scale ultra-short-term prediction data to dynamically correct the day-ahead dispatch plan. The intraday optimization aims to minimize the adjustment amount of controllable power sources in the microgrid and the operating cost of the microgrid. Classic mixed integer programming method and two heuristic algorithms are used respectively, and the results of the three methods are as follows.

[0179] Figure 12 、 Figure 13 and Figure 14 The dynamic correction effect of intraday optimization scheduling on day-ahead plan is shown. By adjusting the controllable unit output, the influence of actual wind and solar output and day-ahead prediction deviation on system dispatch is suppressed, and the specific strategies are as follows:

[0180] 1. Peak period correction: Energy storage discharges in steps and demand response cooperates, cost ratio drops to 68%, and fluctuation rate is ±15%.

[0181] 2. Valley period correction: Energy storage charging power increases to 180 kW, micro-turbine maintains 30 kW, and fuel cost decreases by 12.7%.

[0182] 3. Peak period correction: Micro-turbine over-provisioning and energy storage cooperate to suppress cost surge, and grid fluctuation rate drops by 42%.

[0183] 4. Deep valley period correction: Energy storage SOC increases to 90%, and annual comprehensive cost decreases by 8.7%.

[0184] 5. Flat period correction: Cross-period coordination of energy storage SOC, system inertia increases by 15%.

[0185] The following will compare the day-ahead and intraday optimization scheduling results of the microgrid in this paper. Table 8 gives the controllable power source optimization scheduling results calculated by the classic planning algorithm and two heuristic algorithms in this paper, and the algorithm solving speed comparison, the results are all single objective solving results converted from multi-objective problems by weight.

[0186] Table 8 Comparison of day-ahead and intraday optimization results and algorithm efficiency under three methods

[0187]

[0188] The economic optimization result of the classical mixed integer programming algorithm is better overall

[0189] As can be seen from the comparison of the day-ahead and intraday scheduling optimization results of the classical mixed integer programming and heuristic algorithms in Table 8, in terms of the total economic operation cost, the classical mixed integer programming algorithm is superior to the heuristic algorithm in the day-ahead scheduling and intraday scheduling optimization results and solving efficiency.

Claims

1. A microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis, characterized in that: The following steps are involved: (1) Construct a VMD-LSTM multi-scale prediction framework to improve ultra-short-term accuracy through variable mode decomposition and frequency division prediction; (2) Design a Canopy-spectral clustering-K-means hybrid algorithm to generate typical scenarios based on LHS and enhance the scenario coverage capability; (3) Construct a two-level optimization model, day-ahead and day-intraday, and use a mixed integer programming algorithm to quickly solve high-dimensional problems.

2. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 1 is characterized in that: The step (1) comprises the following steps: (1.1) Extracting multi-scale features of scenery processing through VMD; (1.2) Scaled prediction is performed using LSTM. The prediction process consists of three stages: forgetting stage, selective memory stage, and output stage. (1.3) Construct the VMD-LSTM algorithm process to form a multi-scale prediction framework.

3. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 2 is characterized in that: The VMD used in step (1.1) constructs a constrained variational problem to adaptively decompose the original output sequence f(t) into K intrinsic mode functions (IMFs). Its mathematical expression is: The constraints are: Where u k (t) is the kth IMF component, representing different frequency domain characteristics; ω k is the center frequency of the corresponding component; f(t) is the original wind and solar power output sequence; represents the time derivative, δ(t) is the Dirac function, and j is the imaginary unit; Through iterative solution using the alternating direction multiplier method (ADMM), VMD can effectively eliminate modal aliasing and extract multi-scale features of wind and solar power output.

4. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 2 is characterized in that: The algorithm flow in step (1.3) includes: (1.3.1) Use the VMD algorithm to process the wind and solar power output data under the decomposed time series to obtain the modal components IMFn with different fluctuation characteristics; (1.3.2) Use LSTM neural network to predict the future trend of each modal component; (1.3.3) The prediction components of each modal are added together to calculate the final wind and solar power prediction data.

5. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 1 is characterized in that: The step (2) comprises the following steps: (2.1) Use LHS for scene generation; (2.2) Improve the Canopy-spectral clustering-K-means hybrid algorithm to reduce the scene dimension and extract typical features.

6. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 5 is characterized in that: In step (2.1), LHS is used to generate scenarios. 1000 random scenarios are generated in the probability distribution space of wind and solar power output. LHS ensures that the samples evenly cover all dimensions through stratified sampling. Its mathematical expression is: The probability distribution function F(x) of wind and solar power output is layered and divided into N=1000 equal probability intervals; Randomly select a sample point x in each interval i ,satisfy: Where, F -1 is the inverse function of the cumulative distribution function, ξ i is a uniformly distributed random number; Among them, the fluctuation range of wind power output in the day-ahead scenario is ±30%, and that of photovoltaic power is ±25%, covering typical weather patterns (sunny, cloudy, and overcast); the intraday scenario uses a step size of 15 minutes to simulate ultra-short-term prediction deviations and reflect the rapid fluctuation characteristics of wind and solar power.

7. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 5 is characterized in that: The process of improving the Canopy-spectral clustering-K-means hybrid algorithm in step (2.2) includes the following steps: (2.2.1) Determine the optimal number of scenes K through Canopy coarse clustering; (2.2.2) Use spectral clustering to construct a similarity matrix and reduce the dimension to generate a low-dimensional space matrix Y; (2.2.3) K-means is used to precisely cluster the row vectors of the Y matrix, integrating the advantages of the three algorithms to output K scenarios that represent typical fluctuation characteristics.

8. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 1 is characterized in that: The step (3) comprises the following steps: (3.1) Constructing a two-level optimization framework for day-ahead and intraday; (3.2) Establishment of multi-objective model and constraints.

9. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 8, characterized in that: The optimization framework process in step (3.1) includes: (3.1.1) Generate short-term wind and solar power output forecast curves based on variational mode decomposition (VMD). A random scenario set is constructed through Latin hypercube sampling. A hybrid Canopy-spectral clustering-K-means algorithm is used to reduce the scenarios and extract typical output patterns. (3.1.2) Establish a 24-hour scheduling model, using typical scenarios as input and optimizing the interactive power plan of micro turbines, energy storage, and grid with a 1-hour time resolution, to find the optimal economic-environmental solution for the entire day. (3.1.3) Based on the 15-minute ultra-short-term forecast data within the day, a rolling optimization model is constructed, and the dispatch instructions are updated every hour to dynamically correct the power gap caused by the deviation of wind and solar power forecasts. Mixed integer programming (MILP) is used to achieve minute-level real-time response, ensuring the coordinated optimization of day-ahead planning and real-time operation.

10. The microgrid day-ahead and intraday optimization scheduling method based on multi-time scale scenario analysis according to claim 8, characterized in that: The multi-objective model in step (3.2) is divided into a day-ahead multi-objective model and an intraday dynamic correction model, and the constraints include power balance constraints and an energy storage dynamic model; The economic objectives of the multi-objective model in the day-ahead phase are: Where, T=24 is the total number of scheduling periods; C WT-PV,t is the comprehensive cost of wind and solar power generation in period t (including penalty for energy abandonment and operation and maintenance costs, unit: yuan); C MT,t is the fuel and maintenance cost of the micro-turbine; C ESS,t is the energy storage charging and discharging loss cost; C Grid,t is the grid interaction cost (positive for electricity purchase and negative for electricity sales); The environmental benefit objectives of the multi-objective model in the day-ahead phase are: Where j = 3 is the pollutant type (CO2, SO2, NOx); λ j is the cost of treating the jth type of pollutant (unit: yuan / kg); E j,t is the emission of the micro-turbine in period t (kg), which is determined by the output P MT,t Calculation of emission factors; The intraday dynamic correction model: minF3=αΔC+(1-α)(|ΔP Grid |+|ΔP MT |+|ΔP ESS |) Where, α = 0.7 is the cost deviation weight; ΔC is the deviation between the real-time operating cost and the day-ahead plan (yuan); ΔP Grid , ΔP MT , ΔP ESS are the grid interaction power, micro-turbine output and energy storage adjustment (kW); The power balance constraint: P WT,t +P PV,t +P MT,t +P Grid,t +P ESS,t =P Load,t Where, P WT,t 、P PV,t P is the wind power and photovoltaic power output in period t (kW); MT,t is the output of the micro-turbine (kW); P Grid,t is the grid interaction power (positive for power purchase and negative for power sales); P ESS,t is the energy storage charging and discharging power (discharging is positive, charging is negative); P Load,t is the load demand (kW); The energy storage dynamic model: Where, SOC t is the energy storage charge state during period t; η ch ,η dis is the charge and discharge efficiency; E ESS is the rated capacity of energy storage; Δt is the time interval.

Images