Two-stage low-carbon scheduling method based on carbon entropy and LSTM load prediction

Abstract

The invention discloses a two-stage low-carbon scheduling method based on carbon entropy and LSTM load prediction, and relates to the technical field of power system low-carbon optimization, comprising a short-term load prediction model, a source side optimization scheduling model, a carbon entropy model, a load side demand response model and a two-stage double-layer optimization model. The short-term load prediction model outputs a load curve in the next 24 hours as the input of the source-side optimization scheduling model, the output of each unit calculated by the source-side optimization scheduling model is used as the basis of the carbon entropy model, and the carbon entropy model transmits the calculated node carbon potential to the load-side demand response model; and finally, carrying out iterative optimization on the two-stage double-layer optimization model until convergence. According to the invention, by establishing the source-side optimization scheduling model and the load-side demand response model, collaborative interaction, global optimization and low-carbon economy of source-side and load-side resources are realized, wind curtailment and light curtailment are reduced, the calling demand of a high-carbon unit is reduced from the source, and the total operation cost and the total carbon emission amount of the system are greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of low-carbon optimization technology for power systems, specifically a two-stage low-carbon dispatching method based on carbon entropy and LSTM load forecasting. Background Technology

[0002] As the power system rapidly develops towards a high proportion of renewable energy integration, the distribution network, as a key link connecting power sources and users, is facing severe challenges in its low-carbon transformation. Existing dispatching methods mostly focus on the optimization and regulation of the generation side, failing to fully stimulate and utilize the demand response potential of the load side. At the same time, traditional carbon accounting methods for emission factors cannot accurately track the dynamic distribution of carbon emissions along with power transmission in the network and the differences at the node level, making it difficult to formulate and implement refined carbon emission reduction strategies and fully tap the overall carbon emission reduction potential of the system. Moreover, existing methods mostly adopt a single-stage optimization model, lacking an effective information interaction and collaborative feedback mechanism between the source and load sides. This results in generation planning and load adjustment often being carried out in isolation, failing to form a linkage optimization between the source and load sides, making it difficult for demand response to effectively play its role in reducing the carbon cost of the system.

[0003] Patent CN113762650B discloses an optimization method and system for a distributed predictive power grid. The patent enables optimized scheduling of all distributed power sources in the distribution network, allowing the distribution network to meet both reliability and low-carbon requirements.

[0004] The aforementioned patent dynamically adjusts distributed power sources until the neural network obtains a predicted value of the contribution at a future time that is consistent with the adjustment benchmark, so as to enable the distribution network to meet both reliability and low-carbon requirements. However, there is still room for optimization in terms of demand response on the load side.

[0005] To this end, this application proposes a two-stage low-carbon dispatching method for distribution networks based on carbon entropy and LSTM load forecasting, which involves iterative optimization at both the source and load sides. Summary of the Invention

[0006] The purpose of this invention is to provide a two-stage low-carbon dispatching method based on carbon entropy and LSTM load forecasting, in order to solve the technical problems mentioned in the background art, which are that the existing technology mainly relies on generation-side regulation and fails to effectively coordinate the load-side response potential, traditional carbon accounting methods cannot accurately track the dynamic distribution of node-level carbon emissions, short-term load forecasting accuracy is insufficient and affects dispatching effectiveness, and single-stage optimization mode lacks a source-load interaction feedback mechanism, thus resulting in insufficient tapping of the carbon emission reduction potential of the distribution network and difficulty in achieving low-carbon and economic operation goals in a coordinated manner.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a two-stage low-carbon scheduling method based on carbon entropy and LSTM load prediction, the method comprising the following steps: S1. Construct a short-term load forecasting model based on LSTM, input historical load data and output the load curve for the next 24 hours, as input to the source-side optimized scheduling model; S2. Establish the first-stage source-side optimization scheduling model. Based on the predicted load curve, power flow constraints, and output constraints of each generator set, minimize the system operation calculation to obtain the output of each generator set. S3. Establish a carbon entropy model based on current tracking. By analyzing the current path from the source-side unit output to the load node, quantify the carbon emission contribution of each power source to the load. At the same time, obtain the node carbon potential through the unit output input of the source-side optimized scheduling model. S4. Establish a second-stage load-side demand response model to guide load adjustment direction through nodal carbon potential and reduce power generation costs; S5. Establish a two-stage, two-layer optimization model, feed back the updated load data generated by the load-side demand response model to the source-side optimization scheduling model, and perform iterative optimization until convergence.

[0008] Preferably, the LSTM-based short-term load forecasting model is trained on historical load data through a long short-term memory network to output a load curve for the next 24 hours. The load curve represents the trend of active load over time and is used to provide accurate load input for the source-side optimization scheduling model.

[0009] Preferably, the current-tracking-based carbon entropy model establishes a direct carbon emission contribution mapping between power sources and loads by analyzing the current path from the source-side unit output to the load node. Analyzing the current path involves tracing the current flow path from each source unit to each load node in the power grid topology and calculating the current distribution of each path. The carbon emission contribution mapping is based on the current distribution to form the carbon emission contribution relationship between the power source and the load.

[0010] Preferably, the first-stage source-side optimization scheduling model optimizes the power generation plans of high-carbon and low-carbon units, the actual output of photovoltaic and wind power, and the power purchase plan from the upper-level grid based on the predicted load curve, in order to reduce the total operating cost of the system. At the same time, it obtains the real-time energy transmission structure of the system through power flow calculation, in order to coordinate and optimize power generation resources and power purchase strategies.

[0011] Preferably, in the second stage, the load-side demand response model adjusts the distribution and size of transferable and attenuable loads based on nodal carbon potential and time-of-use pricing signals, and then dynamically optimizes electricity consumption periods and power through the demand response mechanism to reduce the total electricity cost on the load side. The nodal carbon potential is obtained from the calculation results of the carbon entropy model, and the time-of-use pricing signals are obtained from external electricity market data.

[0012] Preferably, the two-stage bi-layer optimization model transfers the node carbon potential calculated by the carbon entropy model to the load-side demand response model in the second stage, and feeds back the optimized load data in the second stage to the source-side optimization scheduling model in the first stage as the initial condition for a new round of iterative optimization.

[0013] Preferably, the total operating cost of the system includes power generation cost, grid purchase cost, wind and solar curtailment penalty cost, and system carbon cost. The system carbon cost adopts a tiered carbon pricing mechanism, which divides the carbon emission into tiered intervals, with each interval corresponding to a different carbon price increment coefficient, to quantify the carbon emission cost.

[0014] Preferably, the source-side optimized scheduling model introduces operational constraints for the energy storage device. These constraints include upper and lower limits of the energy storage state of charge, maximum charge and discharge power limits, charge and discharge efficiency parameters, and charge and discharge power change rate limits, to ensure the safe and stable operation of the energy storage system within the scheduling cycle.

[0015] Preferably, the power flow calculation solves for the voltage magnitude and phase angle of each node by establishing active power balance equations, reactive power balance equations and system operation constraints. The power flow calculation also satisfies the bus voltage safety range and branch transmission capacity limits.

[0016] Preferably, the iterative optimization includes a convergence determination condition. By comparing the change in load data between two iterations with a preset tolerance threshold, if the change is less than the preset tolerance threshold, the optimization result is determined to have converged and the iteration is terminated.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention establishes a source-side optimized scheduling model and a load-side demand response model, realizing the coordinated interaction and global optimization of source-side and load-side resources. It solves the problems of source-load separation, insufficient low-carbon potential tapping, and difficulty in fully utilizing the load-side regulation capacity to reduce costs and carbon emissions in traditional scheduling. The source-side optimized scheduling model forms an economical and low-carbon power generation plan based on power generation cost, power purchase cost, wind and solar curtailment penalty cost, and tiered carbon cost. The load-side model actively adjusts electricity consumption behavior based on the carbon potential signal output from the source side, constructing a two-way regulation mode between the source and load sides. It takes into account the low-carbon economic characteristics of various resources, reduces wind and solar curtailment, and reduces the demand for high-carbon units from the source by guiding the load to low-carbon periods, thus achieving a coordinated reduction in total system operating costs and total carbon emissions. 2. This invention constructs a short-term load forecasting model based on LSTM, achieving high-precision and adaptive prediction of future load curves. It solves the problem of low reliability and inaccurate optimization results in source-side optimization scheduling models caused by large load forecasting deviations. The LSTM short-term load forecasting model trains on historical load data using a long short-term memory network deep learning method, capturing long-term load dependencies and outputting high-precision load curves for the next 24 hours. This solves the problem of low reliability in source-side optimization scheduling models due to inaccurate input data, providing a reliable data foundation for subsequent optimization. It significantly improves the accuracy of scheduling input information, making the power generation plans and power purchase strategies formulated by the source-side optimization model closer to actual needs, reducing inaccuracies in optimization results caused by prediction errors. Thus, while ensuring system safety, it further explores economic optimization space and provides important guarantees for low-carbon scheduling. 3. This invention establishes a carbon entropy model based on current tracking, enabling precise source tracing and dynamic quantification of carbon emissions at distribution network nodes. It addresses the shortcomings of traditional average carbon intensity methods, which cannot characterize the spatiotemporal distribution differences of carbon emissions, support precise carbon reduction decisions, or distinguish the true carbon cost of electricity consumption at different nodes and time periods. This results in the load-side demand response model's inability to accurately optimize power generation plans. The carbon entropy model establishes the physical connection between power sources and load nodes through current tracking technology, displaying the source and destination of each kilowatt-hour of electricity. This allows for the calculation of the real-time node carbon potential of each load node, enabling the source-side optimization scheduling model to identify the nodes and time periods with the highest carbon intensity. This allows for more rational scheduling of unit start-up and shutdown. Simultaneously, by guiding the load-side demand response model, it prioritizes reducing the load on high-carbon-potential nodes and shifting it to low-carbon-potential time periods, suppressing the flow of high-carbon electricity at its source and improving the targeting and effectiveness of carbon reduction strategies. 4. This invention establishes a two-stage, two-layer optimization model, achieving closed-loop iteration and collaborative convergence between the source-side optimization scheduling model and the load-side demand response model. This solves the problems of existing single-stage optimization models, such as the lack of information interaction and collaborative feedback between the source and load sides, the difficulty in fully utilizing demand response potential, and the potential for local optima in the optimization results of the single-stage source-side optimization scheduling model to fail to achieve optimal global resource allocation. In each iteration, the node carbon potential output by the carbon entropy model guides load transfer, and the new load curve formed by the load-side demand response model serves as an updated boundary condition fed back to the source-side optimization scheduling model, triggering a re-optimization of the power generation plan. This cycle repeats until the power generation plan reaches global optima. Through iterative optimization, the final scheduling scheme achieves global coordination and optimality in terms of economy, low carbon emissions, and safety, avoiding the one-sidedness of unilateral optimization. This enables the system to dynamically adapt to load fluctuations, enhancing the robustness and adaptability of distribution network scheduling under high-proportion renewable energy access, and is a crucial guarantee for achieving the coordinated low-carbon goals of the source and load sides. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a flowchart of the LSTM-based short-term load forecasting model of the present invention; Figure 3 The flowchart for establishing the first-stage source-side optimization scheduling model of this invention is shown below; Figure 4 The flowchart for establishing the carbon entropy model based on current tracking in this invention is shown below; Figure 5 The flowchart for establishing the second-stage load-side demand response model of this invention is shown below. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 and Figure 2 The present invention provides an embodiment of a two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting. The LSTM-based short-term load forecasting model is trained on historical load data through a long short-term memory network and outputs a load curve for the next 24 hours. The load curve is the trend of active load change over time and is used to provide accurate load input for the source-side optimization scheduling model. Furthermore, historical load data is collected from the power distribution network data acquisition and monitoring system. The load data includes hourly power load values ​​over the past year. After obtaining the load data, data cleaning is performed to check and process missing and outlier values. For missing data in individual hours, linear interpolation is used to fill in the missing values. Linear interpolation uses the average of the data from two consecutive time points as the missing values. For outliers that deviate significantly from the normal range, Z-score standardization is used for identification and smoothing. The Z-score method determines whether a data point is an outlier by calculating the standard deviation multiple of the data point from the overall mean. Then, the cleaned data is divided into three mutually exclusive sets: training set, validation set, and test set. 70% of the data is used as the training set for learning model parameters, 20% is used as the validation set to adjust model hyperparameters and monitor for overfitting during training, and the remaining 10% is used as the test set to finally evaluate the generalization performance of the model. Then, a short-term load forecasting model based on LSTM is constructed. The short-term load forecasting model is divided into an input layer, an LSTM layer, and a fully connected layer. The input layer is used to receive three-dimensional data input. The LSTM layer is a stacked LSTM layer, with each LSTM layer containing 128 neurons. After the LSTM layer, a fully connected layer is connected. The fully connected layer has 24 neurons. The fully connected layer is used to map the high-dimensional temporal features learned by the LSTM layer to the final prediction output. The prediction output is the load value for the next 24 hours. Then, the short-term load forecasting model is trained using the prepared training and validation sets. The training set data is input into the model, and the number of training cycles is set to 100. During training, the model's performance is evaluated on the validation set simultaneously. If the validation set data is normal for 10 consecutive training cycles, training is automatically stopped. After training, the model's generalization ability is finally evaluated using a test set that has never participated in training and tuning. The mean absolute percentage error (MAPE) of the model on the test set is calculated. If the MAPE is less than 5%, the model passes the accuracy test. The LSTM model that passes the test is then integrated into the distribution network's low-carbon dispatching system to achieve automated operation of short-term load forecasting. The 24-hour load curve output by the LSTM short-term load forecasting model is used as input data and directly passed to the subsequent first-stage source-side optimization dispatching model, thereby initiating the two-stage low-carbon dispatching process of the entire distribution network. This provides a reliable data foundation for the entire dispatching system and significantly improves the accuracy and effectiveness of optimization dispatching decisions.

[0021] Please see Figure 1 and Figure 3 This invention provides an embodiment of a two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting. In the first stage, the source-side optimization scheduling model optimizes the power generation plans of high-carbon and low-carbon units, the actual output of photovoltaic and wind power, and the power purchase plan from the upper-level grid based on the predicted load curve. At the same time, it obtains the real-time energy transmission structure of the system through power flow calculation to coordinate and optimize power generation resources and power purchase strategies. The total operating cost of the system includes power generation cost, grid power purchase cost, wind and solar curtailment penalty cost, and system carbon cost. The system carbon cost adopts a tiered carbon price mechanism, which is divided into multiple tiered intervals based on carbon emissions. Each interval corresponds to a different carbon price increment coefficient. The source-side optimization scheduling model introduces the operation constraints of energy storage devices, including the upper and lower limits of the energy storage state of charge, the maximum charging and discharging power limit, the charging and discharging efficiency parameter, and the charging and discharging power change rate limit. The power flow calculation solves the voltage amplitude and phase angle of each node by establishing the active power balance equation, reactive power balance equation, and system operation constraints. The power flow calculation also satisfies the bus voltage safety range and branch transmission capacity limit. Furthermore, we first establish the objective function. The expression for the source-side optimization scheduling model in the first stage is as follows:

[0022] C total C represents the total cost of system operation. trace C is the cost of purchasing electricity from the grid. diss To mitigate the penalties for curtailing wind and solar power, C gen For the cost of electricity generation, C carbon For the system carbon cost, P H and P L C represents the dispatched power of the high-carbon unit and the low-carbon unit at time t, respectively. H and C L The unit power generation costs for high-carbon and low-carbon units are respectively, d solar and d wind C represents the amount of electricity wasted from solar and wind power, respectively. solar and C wind P represents the corresponding penalty price for curtailing light and wind power, respectively. buy C represents the amount of electricity purchased from the power grid. buy The unit purchase price of electricity is γ, where γ represents the benchmark carbon price and δ represents the tiered carbon price increment coefficient. A The value represents carbon emissions, and 'l' represents the capacity of each step. Meanwhile, the implementation of model optimization must satisfy the physical laws of the power grid and the constraints of safe operation. The physical laws of the power grid are realized through power flow equations, and power flow calculations are based on the node power balance equations.

[0023] After completing the power flow calculation, it is necessary to ensure that all system components operate within safe limits, including maintaining node voltages within permissible ranges and ensuring that line power does not exceed transmission capacity limits.

[0024] P m and Q m These represent the active power and reactive power injected into bus m, respectively, where N represents the number of buses, and V... m and V n θ represents the voltage magnitudes at bus m and bus n. m and θ n G corresponds to the phase angles of bus m and bus n, respectively. mn and B mn V represents the conductance and susceptance of a branch in a network. min and V max P represents the minimum and maximum values ​​of the bus voltage, respectively. mn,max and Q mn,max These represent the maximum active and maximum reactive transmission capabilities of branch mn, respectively. After completing network security constraints, operational constraints are then imposed on power generation resources and energy storage equipment, as well as unit output constraints:

[0025] Renewable energy output constraints:

[0026] External power purchase constraints:

[0027] P High (t) represents the output of the high-carbon unit at time t, P min High and P max High P represents the minimum and maximum allowable output of the high-carbon unit, respectively. Low (t) represents the output of the low-carbon unit at time t, P min Low and P max Low P represents the minimum and maximum allowable output of the low-carbon unit, respectively. solar (t) and P wind (t) represents the photovoltaic and wind power output at time t, P max solar and P max wind P represents the maximum permissible output of photovoltaic and wind power. buy (t) represents the power purchased from the grid at time t, P max buy Indicates the maximum permissible power purchase capacity; The dynamic equation of the state of charge in an energy storage system is:

[0028] The upper and lower limits of SOC are:

[0029] The charging and discharging power is limited to:

[0030] The rate of change of charge / discharge power is limited to:

[0031] E t P represents the state of charge of the stored energy at time t. t ch and P t dis Representing charging power and discharging power respectively, η c and η dThese represent the charging efficiency coefficient and the discharging efficiency coefficient, respectively, where Δt represents the time step, and E... min and E max E represents the lower and upper limits of SOC, respectively. l =E inf and E T =E fin U represents the initial SOC and the final SOC, respectively. t P represents the operating mode indicator variable. ch max and P dis max Δ represents the maximum charging power and the maximum discharging power, respectively. ch and Δ dis These represent the ramp rate limits for charging and discharging, respectively. For example, a distribution network system with 30 nodes includes 2 high-carbon generating units, 3 low-carbon generating units, 10 photovoltaic nodes, and 5 wind power nodes, and is also equipped with 2 energy storage power stations. The P in the high-carbon generating units... min High =20MW, P max High =100MW, unit power generation cost C H =0.5 yuan / kWh, carbon emission factor is 0.8kgCO2 / kWh, P in low-carbon units min Low =10MW, P max Low =80MW, unit power generation cost C L =0.7 yuan / kWh, carbon emission factor is 0.3kgCO2 / kWh, maximum output of photovoltaic power and wind power is 50MW and 60MW respectively, curtailment penalty C solar =0.4 yuan / kWh, wind curtailment penalty C wind =0.3 yuan / kWh, electricity purchase parameter P max buy =150MW, C buy =0.6 yuan / kWh, carbon price parameters γ=200 yuan / ton, δ=0.2, L=100 tons, energy storage parameters E min =20MWh, E max =100MWh, P ch max =P dis max =30MW, η c =η d=0.95. The total system cost, calculated using the source-side optimization scheduling model, is 1,256,800 yuan. Generation cost accounts for 58%, electricity purchase cost for 22%, carbon cost for 15%, and wind and solar curtailment penalty cost for 5%. The total carbon emissions are 285 tons, while the total carbon emissions of the traditional scheduling method are 365 tons, representing a 22% reduction in total carbon emissions. The node carbon potential distribution is between 0.25-0.65 kgCO2 / kWh. The energy storage system charges during off-peak hours and discharges during peak hours, effectively smoothing the load curve and achieving optimal allocation of distribution network resources. Compared with the traditional scheduling method, the total system cost is reduced by 12% and carbon emissions by 22%, providing effective technical support for the low-carbon operation of the distribution network.

[0032] Please see Figure 1 and Figure 4 The present invention provides an embodiment of a two-stage low-carbon dispatching method based on carbon entropy and LSTM load forecasting. The carbon entropy model based on current tracking establishes a direct carbon emission contribution mapping between power sources and charges by analyzing the current path from the source node to the load node. Analyzing the current path involves tracing the current flow path from each source-side unit to each load node in the power grid topology and calculating the current distribution of each path. The carbon emission contribution mapping is based on the current distribution to form the carbon emission contribution relationship between the power source and the load. Furthermore, the change in carbon entropy in the power system is contributed by both carbon entropy flow and carbon entropy production. The carbon entropy model determines the carbon entropy transfer relationship between source and load based on the mapping relationship of energy flow from a single power source to a single electrical load. Finally, through the superposition property of carbon entropy, the carbon entropy of each power source is summed to obtain the total carbon entropy of a single electrical load. The calculation process of the carbon entropy model is to determine the current contribution of the source-side units to each node in the system through the current tracking method. First, the current contribution distribution of each node in the system under the individual action of the generator is tracked, and the formula is:

[0033] This represents the contribution of node current from the action of generator k alone. Let B represent the current phasor flowing from generator k, B be the current correlation coefficient, and n be the number of nodes in the network. This represents the set of branches that flow into node j in the power system. Then, based on the current contribution obtained from tracking, the active power delivered by generator k to node j is calculated. The expression for the active power absorbed by node j from generator k is:

[0034] express The conjugate of the j-th element, Let J represent the voltage vector at node j. The carbon entropy in the formula that represents the transfer of active power is the carbon entropy flow. Next, calculate the active power loss. The expression for the active power loss from generator k to node j is:

[0035] Let be the set of all current flow paths from power node k to node j. This represents the current conjugate transmitted from generator k to node j along path l. This indicates the voltage drop across branch α. This represents the voltage drop between generator k and node j. The carbon entropy transferred by the active power loss calculated in the formula is the carbon entropy production. Then, the carbon entropy transferred from generator k to node j is calculated, expressed as:

[0036] This represents the carbon potential of generator k; Finally, the nodal carbon potential is calculated, and the expression is:

[0037] This represents the carbon potential at node j. This represents the set of nodes where all generators are located in the power system. This represents the active load of node j, and finally the output node carbon potential value is transmitted to the second-stage model.

[0038] Please see Figure 1 and Figure 5 The present invention provides an embodiment of a two-stage low-carbon dispatching method based on carbon entropy and LSTM load forecasting. In the second stage, the load-side demand response model adjusts the distribution and size of transferable and attenuable loads according to the nodal carbon potential and time-of-use electricity price signal. Then, the demand response mechanism dynamically optimizes the electricity consumption period and power to reduce the total electricity cost on the load side. The nodal carbon potential is obtained from the calculation results of the carbon entropy model, and the time-of-use electricity price signal is obtained from external electricity market data. Furthermore, the carbon potential of a node is calculated by the carbon entropy model. The higher the carbon potential, the higher the environmental cost caused by the electricity consumption behavior at that node. Time-of-use pricing is determined by the electricity market, with different prices applied at different times of the day. Peak hours have higher prices, while off-peak hours have lower prices. Transferable loads are loads that can be shifted at different times of the day without affecting the total electricity consumption and service functions. Reduceable loads are loads whose power is reduced or interrupted within a certain range during a specific period of time, achieved by sacrificing a small amount of comfort or adjusting the operating mode. The direct objective of the second-stage load-side demand response model is to minimize the total electricity cost on the load side during the scheduling cycle. The objective function expression is:

[0039] σ t P represents the unit electricity price at time t, which is the sum of the unit electricity price and the unit carbon price. z,t L Let T represent the active load of node z at time t, where T is the total number of time periods in a scheduling cycle, such as 24 hours. Then, constraints are applied to the load adjustment range, for example, if node z has a transferable load P at time t. z,t L,shift Adjustments must meet time-period load transfer limits. These limits stipulate that the load transfer amount within any given time period must not exceed a certain percentage (e.g., 20%) of the original baseline load to prevent instantaneous impacts on the power grid. The constraint formula is: -0.2P z,t L,base ≤P z,t L,shift ≤0.2P z,t L,base ; Simultaneously, the algebraic sum of the transfer volume across all time periods within the complete scheduling cycle is zero, ensuring that the user's total electricity demand is met, which can reduce the load P. z,t L,curt The adjusted total node load P cannot exceed 30% of the original baseline load. z,t L =P z,t L,base +P z,t L,shift -P z,t L,curt ; Then input the carbon potential and time-of-use electricity price σ for each node and time period provided by the first stage and the carbon entropy model. t and initial reference load P z,t L,base The second-stage load-side demand response model then outputs the final optimized load value P for each node in each time period. z,t L The specific transfer amount P of each node in each time period z,t L,shift and reduction amount P z,t L,curt Then update the load data P z,t L Feedback is sent back to the first-stage source-side optimized scheduling model for a new round of iteration; For example, the evening peak hours for a residential area are 18:00-19:00 (time period t1), and the late-night hours are 02:00-03:00 (time period t2). The input baseline load P... t1 L,base =100kW, P t2 L,base =30kW, time-of-use electricity price σ t1 =1.2 yuan / kWh, σ t2 =0.4 yuan / kWh, node carbon potential is Φ t1 =0.5kgCO2 / kWh, Φ t2 =0.1kgCO2 / kWh, the nodal carbon potential is calculated by the carbon entropy model, the transferable load is 20kW, data shows that the off-peak electricity price is only 1 / 3 of the peak electricity price, and the carbon potential of time period t2 is only 1 / 5 of that of time period t1. Transferring load can significantly reduce electricity costs and implicit carbon costs. Then, in the second stage, the load-side demand response model transfers 15kW of transferable load from time period t1 to time period t2. The adjusted load of time period t1 is P. z,t L =100-15=85kW, the adjusted load for time period t2 is P z,t L =30+15=45kW; The original hourly electricity cost was 15kW × 1h × 1.2 = 18 yuan, while the adjusted hourly cost is 15kW × 1h × 0.4 = 6 yuan, resulting in a saving of 12 yuan. Simultaneously, the carbon emissions before the transfer were 15kWh × 0.5kg / kWh = 7.5kgCO2, while after the transfer, the carbon emissions are 15kWh × 0.1kg / kWh = 1.5kgCO2, a reduction of 6kgCO2 emissions. Furthermore, guiding the load transfer in time and space brings direct economic benefits to users and indirectly promotes a reduction in the overall use of high-carbon energy by the power grid, achieving a win-win situation for both user economic benefits and the system's low-carbon goals. Moreover, the optimized load curve provides lower-carbon and more economical load inputs for further optimization of the first-stage model, thus achieving low-carbon and low-cost dispatch in the two-stage iteration.

[0040] Please see Figure 1 The present invention provides an embodiment of a two-stage low-carbon scheduling method based on carbon entropy and LSTM load prediction. The two-stage dual-layer optimization model transfers the node carbon potential calculated by the carbon entropy model to the load-side demand response model in the second stage, and feeds back the optimized load data in the second stage to the source-side optimization scheduling model in the first stage as the initial condition for a new round of iterative optimization. The iterative optimization is set with a convergence judgment condition. By comparing the change in load data in the two iterations with a preset tolerance threshold, if the change is less than the preset tolerance threshold, the optimization result is determined to be converged and the iteration is terminated. Furthermore, the two-stage, two-layer optimization model places the source-side optimization scheduling model at the upper layer and the load-side demand response model at the lower layer. The source-side optimization scheduling model is connected to the load-side demand response model through the carbon entropy model. First, the load curve predicted by LSTM is used as the input to the first-stage source-side optimization scheduling model to calculate the output of each unit and the scheduling plan. Then, the first-stage source-side optimization scheduling model is run, and the unit output is input into the carbon entropy model to obtain the nodal carbon potential Φ. l Then Φ l The data is transferred to the second-stage load-side demand response model, where demand response optimization is performed to obtain the updated load data P. z,t L , will P z,t L Feedback is sent back to the first-stage model to update the load input. The two-stage, two-layer optimization model calculates the change in the updated load, reads the load data generated in the current iteration k, and then reads the load data generated in the previous iteration k-1. These two sets of data are then substituted into a preset variable formula for calculation. The variable formula is as follows:

[0041] ΔP (k) N represents the total load change in the k-th iteration. node Let P be the total number of load nodes, and T be the total number of scheduling periods, such as 24 hours. If ΔP (k) If ≤ ε, the system determines that the optimization result has converged and outputs the final scheduling scheme; otherwise, ΔP is decremented. (k+1) Feedback is given back to the first-stage model, let k = k + 1, and continue the next iteration until convergence, where ε is the tolerance threshold. For example, after three iterations, ΔP in the power distribution network system (1) =0.05, ΔP (2) =0.01, ΔP (3) =0.002, tolerance threshold ε is set to 0.003, ΔP (3) =0.002 < 0.03, which satisfies the convergence condition, and the iteration terminates. Finally, the system outputs the optimal source-side power generation plan, energy storage strategy, power purchase plan and load adjustment strategy to achieve synergistic minimization of total operating cost and carbon emission cost.

[0042] Working principle: The system first predicts the short-term load based on the LSTM model as the initial input for scheduling. In the first stage, the source-side optimization scheduling model optimizes various power generation resources and power purchase plans. Then, the real-time operating status of the power grid is obtained through power flow calculation. Finally, the carbon potential of each load node is calculated through the carbon entropy model to form the carbon cost spatial distribution. In the second phase, the load-side demand response model dynamically adjusts the spatiotemporal distribution of transferable and reduceable loads through carbon potential signals and time-of-use pricing, guiding users to avoid electricity consumption during high-carbon and high-price periods, reducing local carbon footprint and electricity costs. The optimized load data is fed back to the source-side optimization scheduling model in the first phase, triggering a new round of power generation plan adjustments. The source-side optimal scheduling model and the load-side demand response model are connected by a carbon entropy model. The source-side optimal scheduling model is set at the upper layer of the two-stage dual-layer optimization model, and the load-side demand response model is set at the lower layer of the two-stage dual-layer optimization model. Carbon potential guides load transfer, and load changes, in turn, affect the power generation structure and carbon potential distribution until the system reaches a coordinated optimal state. Under the premise of ensuring grid security, the low-carbon potential of source-side, load-side and energy storage resources is fully explored, and the overall reduction of carbon emissions and operating costs is achieved, providing a feasible technical path for the low-carbon scheduling of new power systems.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting, characterized in that: The method comprises the following steps: S1, constructing an LSTM-based short-term load prediction model, inputting historical load data and outputting a load curve of the next 24 hours as input of a source-side optimal scheduling model; S2, establishing a first-stage source-side optimal scheduling model, minimizing system operation calculation to obtain the output of each generator according to the predicted load curve, power flow constraints and the output constraints of each generator; S3, establishing a carbon entropy model based on current tracking, quantifying the carbon emission contribution of each power source to the load by analyzing the current path from the source-side unit output to the load node, and obtaining the node carbon potential through the unit output input by the source-side optimal scheduling model; S4, establishing a second-stage load-side demand response model, guiding the load adjustment direction through the node carbon potential to reduce the power generation cost; S5, establishing a two-stage double-layer optimization model, feeding back the updated load data generated by the load-side demand response model to the source-side optimal scheduling model for iterative optimization until convergence.

2. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 1, characterized in that: The LSTM-based short-term load prediction model trains the historical load data through a long short-term memory network to output a load curve of the next 24 hours, and the load curve is the trend of active load over time, which is used to provide accurate load input for the source-side optimal scheduling model.

3. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 1, characterized in that: The carbon entropy model based on current tracking establishes a direct carbon emission contribution mapping between power sources and loads by analyzing the current path from the source-side unit output to the load node, and the current path is tracking the current flow path from each source unit to each load node in the grid topology and calculating the current distribution of each path, and the carbon emission contribution mapping is the carbon emission contribution relationship of power sources to loads according to the current distribution.

4. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 1, characterized in that: The first-stage source-side optimal scheduling model optimizes the power generation plan of high-carbon units and low-carbon units, the actual output of photovoltaic and wind power, and the power purchase plan from the upper grid according to the predicted load curve, which is used to reduce the total system operation cost, and the real-time energy transmission structure of the system is obtained through power flow calculation to coordinate and optimize the power generation resources and power purchase strategy.

5. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 1, characterized in that: The second-stage load-side demand response model adjusts the distribution and size of transferable load and weakenable load according to the node carbon potential and time-of-use price signal, and dynamically optimizes the power consumption period and power through the demand response mechanism to reduce the total power consumption cost on the load side, the node carbon potential is obtained from the calculation result of the carbon entropy model, and the time-of-use price signal is obtained from the external power market data.

6. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 1, characterized in that: The two-stage double-layer optimization model transmits the node carbon potential calculated by the carbon entropy model to the second-stage load-side demand response model, and feeds back the load data optimized in the second stage to the first-stage source-side optimal scheduling model as the initial condition of the next iteration optimization.

7. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 4, characterized in that: The total system operation cost includes power generation cost, grid power purchase cost, wind and light abandonment penalty cost and system carbon cost, the system carbon cost adopts a step carbon price mechanism, divides the step interval according to the carbon emission, each interval corresponds to a different carbon price increasing coefficient, which is used to quantify the carbon emission cost.

8. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 5, characterized in that: The operation constraint of the energy storage device is introduced in the source side optimization scheduling model, and the operation constraint includes upper and lower limits of the state of charge of the energy storage, maximum charging and discharging power limits, charging and discharging efficiency parameters and charging and discharging power change rate limits, so as to ensure safe and stable operation of the energy storage system in the scheduling period.

9. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 4, characterized in that: The power flow calculation solves the voltage amplitude and phase angle of each node by establishing the active power balance equation, the reactive power balance equation and the system operation constraint, and at the same time, the power flow calculation also meets the bus voltage safety range and the branch transmission capacity limit.

10. The two-stage low-carbon scheduling method based on carbon entropy and LSTM load forecasting according to claim 6, characterized in that: The convergence judgment condition is set in the iterative optimization, the change amount of the load data in the previous and next two iterations is compared with the preset tolerance threshold, and if the change amount is less than the preset tolerance threshold, it is judged that the optimization result converges and the iteration is terminated.

Images