A method, system, medium and device for robust optimization scheduling of building microgrid

Abstract

The present invention proposes a method, system, medium and device for robust optimization scheduling of a building microgrid, which belongs to the technical field of building microgrid optimization scheduling, including: generating a number of source load samples based on the day-ahead source load prediction value, representing the possible realization value of the source load on the next day, thereby constructing a source load uncertainty set; establishing a day-ahead scheduling layer, including: dividing the source load uncertainty set into a larger part of the source load uncertainty set, a smaller part of the source load uncertainty set and establishing a three-stage day-ahead robust optimization scheduling model; on the basis of the day-ahead scheduling layer, establishing an intraday real-time optimization scheduling layer based on model reinforcement learning, and outputting an intraday real-time optimal scheduling plan.

Description

Technical Field

[0001] The present invention belongs to the technical field of building microgrid optimization scheduling, and in particular relates to a building microgrid robust optimization scheduling method and system. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] At present, the energy structure of buildings is relatively simple, mainly fossil energy. With the low-carbon development of buildings, the energy structure of buildings will become diversified, and renewable energy will gradually dominate. In the process of the continuous and synchronous expansion of renewable energy generation and building electricity consumption, building microgrids came into being.

[0004] To ensure the efficient and safe operation of building microgrids, the key lies in efficient dispatching strategies and sufficient adjustable resources. However, when new energy generation such as photovoltaic power generation and wind power generation accounts for a large proportion of the system, the dispatching of microgrids will encounter some difficulties. First, due to the uncertainty of users' electricity usage habits, electricity demand is random; secondly, the output of new energy generation such as photovoltaic power generation and wind power generation is also affected by natural conditions such as sunshine intensity and wind power, and is unstable.

[0005] At present, artificial intelligence technology still has limitations in accurately predicting the generation and power demand of new energy sources such as photovoltaic power generation and wind power generation, which leads to prediction bias and increases the problem of source-load uncertainty, which has a direct impact on the scheduling plan of microgrids. Secondly, with the large-scale grid connection of new energy sources such as photovoltaic power generation and wind power generation, the capacity of traditional adjustable power generation equipment has been relatively reduced, which makes the adjustable resources used in microgrids to balance the fluctuations of new energy generation such as photovoltaic power generation and wind power generation and power demand insufficient, further exacerbating the impact of the source-load uncertainty problem.

[0006] After searching, it is found that there are relevant technical solutions for building microgrids in the prior art, but the existing technical problems are: the traditional building microgrid optimization scheduling usually includes two levels: one is the two-stage robust optimization scheduling of the day before, and the other is the real-time optimization scheduling within the day. In the day-ahead scheduling, the construction of the source-load uncertainty set does not fully utilize the correlation between the source-load prediction deviation and the predicted value, which makes the constructed uncertainty set too conservative. In addition, the two-stage robust optimization scheduling does not take into account the multi-time scale characteristics of the source-load equipment and uncertainty in the building microgrid, which makes the scheduling plan of the source-load side equipment with short time scale characteristics too conservative. At the level of intraday real-time scheduling, the traditional mathematical programming-based method cannot fully consider the randomness of the source-load prediction deviation, so it is difficult to effectively deal with the impact of the source-load prediction deviation. Summary of the invention

[0007] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a robust optimization scheduling method for a building microgrid. Based on a daily real-time optimization scheduling model based on model reinforcement learning, the day-ahead scheduling results are used to calculate the intraday real-time scheduling results to ensure stable and efficient operation of the building microgrid system.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] In a first aspect, a robust optimization scheduling method for a building microgrid is disclosed, comprising:

[0010] Based on the day-ahead source load forecast value, several source load samples are generated to represent the possible realized value of the source load on the next day, thereby constructing a source load uncertainty set;

[0011] Establishing a day-ahead scheduling layer includes: dividing the source-load uncertainty set into a larger source-load uncertainty set and a smaller source-load uncertainty set, and establishing a day-ahead three-stage robust optimization scheduling model, wherein the day-ahead three-stage robust optimization scheduling model includes:

[0012] Phase 1: Determine the dispatch plan under the nominal scenario of source load forecast value;

[0013] Phase II: To maximize the minimum recourse cost of the building microgrid under the worst scenario of a larger source-load uncertainty concentration, the recourse plan of the gas turbine, energy storage battery, and the HVAC and electric vehicle aggregate loads of the load aggregator with different response speeds is calculated;

[0014] The second-stage recourse plan is superimposed on the nominal plan of the first stage. In the third stage, the recourse plan of the energy storage battery with faster response speed and the aggregate load of the electric vehicle of the load aggregator is calculated with the goal of maximizing the minimum recourse cost of the building microgrid under the worst scenario of uncertain concentration of a smaller part of the source and load;

[0015] On the basis of the day-ahead scheduling layer, an intraday real-time optimization scheduling layer is established based on model reinforcement learning to output the intraday real-time optimal scheduling plan.

[0016] As a further technical solution, the specific process of constructing the source-load uncertainty set is:

[0017] Get the latest day-ahead forecast value;

[0018] Obtain the expected value and covariance matrix of the conditional correlation between the source load prediction value and the true value;

[0019] The conditional correlation between the source load prediction value and the true value is sampled to generate a number of samples, and then a sample of the source load true value is obtained;

[0020] based on The samples obtained by sampling at the moment and the latest day-ahead forecast value establish a source load uncertainty set.

[0021] As a further technical solution, the source load uncertainty set is decomposed into an uncertainty set with a large fluctuation amplitude and long time scale characteristics and an uncertainty set with a small fluctuation amplitude and short time scale characteristics.

[0022] As a further technical solution, the first stage: determine the dispatch plan under the nominal scenario of source-load forecast value, including the optimization objectives, constraints and decision variables of the first stage;

[0023] The constraints in the first stage include power balance constraints, gas turbine operation constraints, grid constraints, energy storage battery operation constraints, and load aggregator constraints.

[0024] As a further technical solution, the optimization target of the second stage is the second stage recourse cost, which specifically includes the gas turbine recourse cost, the energy storage battery recourse cost and the compensation cost of the building microgrid for the load aggregator HVAC aggregate load regulation behavior;

[0025] The new constraints added in the second stage compared to the first stage are the upper and lower limits of the thermal comfort temperature for HVAC users in the building. The corresponding constraints are intended to limit the indoor temperature to the thermal comfort range of the users.

[0026] As a further technical solution, the third stage recourse costs include the energy storage battery recourse costs and the compensation costs of the building microgrid for the load aggregator's electric vehicle aggregate load regulation behavior;

[0027] Compared with the second stage, the third stage adds the upper and lower limits of the aggregate target charging capacity of electric vehicles; The upper and lower limits of the deviation from the target charging amount that an electric vehicle user can tolerate.

[0028] As a further technical solution, a daily real-time optimization scheduling layer is established based on model reinforcement learning, including:

[0029] The intraday real-time optimization scheduling problem is constructed as a Markov decision process, defining states, actions, and rewards;

[0030] Use LSTM to learn the state transition model of the building microgrid system. The input of the state transition model is the state and action, and the output is the state change;

[0031] The actor network that learns deterministic strategies interacts with the learned building microgrid system state transition model in a way that optimizes the finite prediction interval. The accumulated reward within the finite prediction interval obtained by the interaction is used as the update target of the critic network. The actor and critic network model parameters are continuously updated to finally obtain the optimal control strategy.

[0032] After obtaining the daily source-load forecast value at set intervals, the dispatch center will use the actor network optimal control strategy to obtain the daily real-time dispatch plan.

[0033] In a second aspect, a robust optimization dispatching system for a building microgrid is disclosed, comprising:

[0034] The source load uncertainty set construction module is configured to: generate a number of source load samples based on the day-ahead source load prediction value, representing the possible realization value of the source load on the next day, thereby constructing a source load uncertainty set;

[0035] The day-ahead scheduling layer establishment module is configured to: establish the day-ahead scheduling layer, including: dividing the source-load uncertainty set into a larger source-load uncertainty set and a smaller source-load uncertainty set, and establishing a day-ahead three-stage robust optimization scheduling model, wherein the day-ahead three-stage robust optimization scheduling model includes:

[0036] Phase 1: Determine the dispatch plan under the nominal scenario of source load forecast value;

[0037] Phase II: To maximize the minimum recourse cost of the building microgrid under the worst scenario of a larger source-load uncertainty concentration, the recourse plan of the gas turbine, energy storage battery, and the HVAC and electric vehicle aggregate loads of the load aggregator with different response speeds is calculated;

[0038] The second-stage recourse plan is superimposed on the nominal plan of the first stage. In the third stage, the recourse plan of the energy storage battery with faster response speed and the aggregate load of the electric vehicle of the load aggregator is calculated with the goal of maximizing the minimum recourse cost of the building microgrid under the worst scenario of uncertain concentration of a smaller part of the source and load;

[0039] The intraday real-time optimization scheduling module is configured as follows: on the basis of the day-ahead scheduling layer, an intraday real-time optimization scheduling layer is established based on model reinforcement learning, and the output is the intraday real-time optimal scheduling plan.

[0040] One or more of the above technical solutions have the following beneficial effects:

[0041] In order to overcome the limitations of the traditional single-layer deterministic optimization scheduling method, the technical solution of the present invention innovatively proposes a multi-level building microgrid management framework. The framework integrates data-driven uncertainty set modeling technology and a three-stage robust optimization scheduling strategy with multi-time scale characteristics, ensuring comprehensive processing of source and load uncertainties in the day-ahead scheduling stage. At the same time, combined with the intraday real-time scheduling method based on model reinforcement learning, the system's responsiveness to forecast deviations and scheduling flexibility are further improved.

[0042] In order to accurately describe the uncertainty of source load, the technical solution of the present invention proposes a data-driven source load uncertainty set modeling method for the correlation characteristics between the source load prediction deviation and the predicted value, which uses the probability information of the source load historical data to capture the conditional correlation between the prediction deviation and the predicted value, so that the source load uncertainty set has stronger data adaptability to the changes in the source load prediction results and reduces conservatism. Secondly, in the day-ahead scheduling layer, in view of the multi-time scale characteristics of the building microgrid source load equipment and source load uncertainty, the present invention proposes a day-ahead scheduling layer based on three-stage robust optimization scheduling. First, the source load uncertainty set is decomposed into two parts with large fluctuation amplitude and long time scale and small fluctuation amplitude and short time scale. Secondly, in the first stage, the nominal scheduling plan of all source load equipment is calculated. In the second and third scheduling stages, based on the long-time scale and short-time source load uncertainty sets, the recourse scheduling plan of the source load equipment with long-time scale characteristics and short-time scale characteristics is obtained, and finally the day-ahead scheduling is summarized. In the intraday real-time scheduling layer, in view of the random characteristics of the source-load prediction deviation, the present invention proposes an intraday real-time optimization scheduling model based on model reinforcement learning. The intraday real-time scheduling results are calculated by using the day-ahead scheduling results to ensure the stable and efficient operation of the building microgrid system.

[0043] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0045] Figure 1 It is a schematic diagram of the overall process of the method of the embodiment of the present invention;

[0046] Figure 2 The figure is a schematic diagram of a specific process of the method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0048] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0049] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0050] Embodiment 1

[0051] This embodiment discloses a robust optimization scheduling method for a building microgrid, which can effectively deal with the uncertainty of sources and loads in the building microgrid, including:

[0052] Step 1: Establish a source-load uncertainty set with correlation information based on data-driven.

[0053] The probability information of the source load historical data and the corresponding prediction deviation are used to capture the conditional correlation between the prediction deviation and the prediction value. Based on the extracted correlation information and the latest source load prediction value, a source load uncertainty set based on the conditional normal connection method is established, so that the source load uncertainty set has stronger data adaptability to changes in the prediction results.

[0054] In an implementation example, in order to effectively solve the source-load uncertainty problem in the optimal scheduling of building microgrids, it is necessary to adopt a robust optimization scheduling method. In this process, it is first necessary to construct a source-load uncertainty set. However, although traditional methods based on probability statistics and optimization theory take into account the worst-case source-load uncertainty, they are often too conservative and may not be sufficient to meet the flexibility and efficiency requirements of practical applications. Since there is a certain correlation between the source-load prediction deviation and the predicted value, this correlation is manifested as follows: when the predicted power is low, the prediction deviation is small and the distribution is concentrated; when the predicted power is high, the prediction deviation is large and the distribution is dispersed.

[0055] In view of this characteristic, this example proposes a data-driven source load uncertainty set modeling method. This method uses the probability information of source load historical data to capture the conditional correlation between prediction deviation and prediction value, thereby improving the adaptability of the source load uncertainty set to changes in prediction results and reducing conservatism.

[0056] First, collect the historical real value of source load The predicted value obtained by the corresponding deep learning-based source load prediction model , that is, collecting the historical real values ​​and corresponding predicted values ​​of load power values ​​such as photovoltaic power generation, wind power generation, electric vehicle batteries, and HVAC equipment; secondly, according to the normal connection theory, the joint cumulative distribution of source-load prediction deviation and its predicted value is obtained.

[0057] (1)

[0058] in, is the joint cumulative distribution of the source load prediction deviation and its predicted value; , is the source load prediction deviation, where ; and are the intermediate variables of normal connection obtained from the historical real value of source load and the corresponding predicted value; and are the cumulative distribution functions obtained from the historical true values ​​of source load and the corresponding predicted values ​​of source load respectively; is the standard normal cumulative distribution function obtained from the historical true value of source load and the corresponding source load forecast value, is its inverse function; and are the probability density functions obtained from the historical true value of source load and the corresponding source load prediction value respectively; Normal connection intermediate variable and The probability density function of the standard multivariate normal distribution is, is the covariance matrix of the probability density function.

[0059] Among them, source-load data refers to the load power values ​​of photovoltaic power generation, wind power generation, electric vehicle batteries, HVAC equipment, etc., among which photovoltaic power generation and wind power generation are on the source side, and the load power values ​​of electric vehicle batteries, HVAC equipment, etc. are on the load side.

[0060] definition , , , is a vector of normally connected intermediate variables obtained from the historical true values ​​of the source load and the corresponding predicted values, and , that is, the covariance matrix is The normal distribution of .

[0061] (2)

[0062] (3)

[0063] in and are the linear correlation coefficient and Spearman correlation coefficient of the normal connection intermediate variable, which are used to extract the correlation between the historical true value and the predicted value of the source load. and Calculate the same .

[0064] Secondly, the conditional probability relationship between the source load prediction deviation and the source load prediction value is established through the normal connection model.

[0065] (4)

[0066] in, is the conditional probability density function between the source load prediction deviation and the source load prediction value; , describes the distribution characteristics of the source load prediction value; is the conditional distribution of the multivariate normal distribution, satisfying , describes the conditional correlation between the source load prediction value and the true value.

[0067] (5)

[0068] in, is the mean vector, is the covariance matrix, Describe the distribution characteristics between the actual value and the predicted value of the source load, Describe the distribution characteristics of the source load prediction value, Describe the distribution characteristics of the true value of the source charge.

[0069] The conditional correlation between the source load prediction value and the true value is described. Using this conditional correlation, several source load samples can be generated based on the latest day-ahead source load prediction value, representing the possible realization value of the source load on the next day, thereby constructing a source load uncertainty set. The specific process is as follows:

[0070] 1) Get the latest day-ahead source load forecast value and get ;

[0071] 2) Using formula (5) to obtain The expected value and covariance matrix of ;

[0072] 3) Through Sampling to generate several of samples, and then through Get a sample of the true value of the source charge.

[0073] based on Samples obtained by sampling at the moment and the latest day-ahead forecast ( is the day-ahead dispatch period), and establish a source-load uncertainty set.

[0074] (6)

[0075] Among them, the sampled vector The distribution characteristics of the forecast deviation under different day-ahead forecast values ​​are shown. Therefore, the constructed source load uncertainty set has a strong ability to adapt to the data changes of the forecast results and can describe the uncertainty of the source load more accurately. is the fluctuation range of the forecast deviation, is a 0-1 variable, and Together they describe the degree of conservatism of the source-charge uncertainty set.

[0076] Step 2: Establish the day-ahead scheduling layer based on the three-stage robust optimization scheduling.

[0077] First, the source-load uncertainty set is divided into two parts: one with large fluctuation amplitude and long time scale and one with small fluctuation amplitude and short time scale, and the conservative type is reduced. Among them, the large and small fluctuation amplitude refers to the large fluctuation amplitude of the prediction deviation, and the long and short time scale refers to the long and short time scale characteristics of the source-load side equipment. Secondly, a three-stage day-ahead robust optimization scheduling model is established. In the first stage, the nominal scheduling plan for all source-load equipment is calculated with the goal of minimizing the operating cost of the building microgrid under the source-load prediction value; the second stage is based on the nominal scheduling plan, with the goal of maximizing the minimum cost of the building microgrid in the source-load uncertainty set with large fluctuation amplitude and long time scale, and the recourse scheduling plan for the source-load equipment with long time scale characteristics is obtained; the third stage is based on the nominal scheduling plan and the recourse scheduling plan of the second stage, with the goal of maximizing the minimum cost of the building microgrid in the source-load uncertainty set with small fluctuation amplitude and short time scale, and the recourse scheduling plan for the source-load equipment with short time scale characteristics is obtained. Finally, the scheduling plans of the three stages are summarized as the day-ahead scheduling plan and sent to the lower-level intraday real-time optimization scheduling layer.

[0078] In the multi-time-scale coordinated day-ahead and intra-day real-time two-layer building microgrid management framework, the robust optimization scheduling method is applied in the day-ahead stage to pre-process possible source-load fluctuations. In the day-ahead scheduling layer, the traditional day-ahead two-stage robust optimization scheduling cannot effectively utilize the multi-time-scale characteristics of the source-load side equipment and source-load uncertainty of the building microgrid. Based on this characteristic, the present invention establishes a day-ahead scheduling layer based on three-stage robust optimization scheduling. First, the source-load uncertainty set is divided into two parts, one with large fluctuation amplitude and long time scale, and the other with small fluctuation amplitude and short time scale, to reduce the conservative type. Secondly, a three-stage day-ahead robust optimization scheduling model is established.

[0079] In the building microgrid, the response speed of the source-load side equipment shows obvious multi-time scale characteristics. For example, due to the limitations of thermal inertia and regulation process, the response speed of gas turbines and HVAC equipment is relatively slow and has long time scale characteristics. At the same time, energy storage batteries, electric vehicle batteries, photovoltaic cells, and wind turbines show fast response capabilities, and can quickly adapt to load changes regardless of low or high power output, and have short time scale characteristics. Considering that the accuracy of source-load prediction is inversely proportional to the prediction time scale, that is, the longer the prediction time, the greater the prediction deviation; the shorter the prediction time, the smaller the deviation, the management strategy of the building microgrid needs to adjust the processing of source-load prediction deviation according to the response speed of the equipment. Specifically, for devices with slower response speed, strategies can be designed to handle larger prediction deviations; while for devices with faster response speed, they can respond quickly and adjust to adapt to smaller prediction deviations. Based on this, the present invention further subdivides the source-load uncertainty set into two parts: one part is the case of large fluctuation amplitude and long time scale, and the other part is the case of small fluctuation amplitude and short time scale. This subdivision helps to manage the uncertainty of source-load side equipment more accurately and optimize the overall dispatch strategy of the building microgrid. Equations (7) and (8) are the larger and smaller source-load uncertainty sets, respectively.

[0080] (7)

[0081] (8)

[0082] , are the larger and smaller source load uncertainty sets, respectively; and are the larger and smaller source load uncertainty concentrations, respectively. The source-load value vector at the moment, i.e., the load power value vector of photovoltaic power generation, wind power generation, electric vehicle batteries, HVAC equipment, etc.; for The source load prediction value vector at the moment, that is, the load power prediction value vector of photovoltaic power generation, wind power generation, electric vehicle batteries, HVAC equipment, etc.; are used to describe the fluctuation range of source load prediction deviation in larger and smaller source load uncertainty sets, respectively, and the value range is ; is a 0-1 variable, used to determine the uncertainty set in the larger and smaller source load uncertainty sets. Whether there is a positive or negative source load prediction deviation at the moment; are the uncertainty parameters in the larger and smaller source load uncertainty sets, representing the total number of time periods where source load prediction deviations occur, and their value range is , ; and Together they describe the degree of conservatism of the source-load uncertainty set. The larger the value, the worse the source-load uncertainty scenario, making the scheduling plan more conservative.

[0083] The optimization objective of the three-stage day-ahead robust optimization scheduling model is given by the following equation (9). In the first stage, we determine the scheduling plan under the nominal scenario of source-load forecast values, and its optimization objective, constraints, and decision variables are defined by the following equations (10)-(12), respectively.

[0084] (9)

[0085] (10)

[0086] (11)

[0087] (12)

[0088] in, Optimize the goals for the first, second and third stages. are the decision vectors for the first, second and third stages, The second and third stages are the larger and smaller source load uncertainty sets. The source charge value vector in .

[0089] Phase I operating costs Including gas turbine start-up and shutdown costs, gas turbine operation and maintenance costs, energy storage battery operation and maintenance costs and electricity purchase costs, is the cost of a single start-up and shutdown of the gas turbine, and Unit operation and maintenance costs for gas turbines and energy storage batteries, for The power selling price of the power grid at that moment, For the first stage The gas turbine start and stop status at each moment. For the first stage The gas turbine output power at time For the first stage The charging and discharging power of the energy storage battery at all times, For the first stage The power purchased from the grid at all times; It is the day-ahead scheduling time scale.

[0090] The constraints in the first stage include power balance constraints, gas turbine operation constraints, grid constraints, energy storage battery operation constraints, and load aggregator constraints. The above formula (11) includes power balance constraints and gas turbine operation constraints. For the first stage Photovoltaic power output at all times, For the first stage Wind power output power at any moment, For the first stage Load power consumption at all times; For the first stage The total load of HVAC and electric vehicles of the load aggregator at the moment. Equation (12) is the decision vector of the first stage.

[0091] In the second stage, the goal is to maximize the minimum recovery cost of the building microgrid under the worst scenario of a large part of the source and load uncertainty concentration, and calculate the recovery plan of the gas turbine, energy storage battery, and the load aggregator HVAC and electric vehicle lumped load with different response speeds. The following equations (13)-(15) are the optimization objectives of the second stage, the newly added constraints and decision variables relative to the first stage.

[0092] (13)

[0093] (14)

[0094] (15)

[0095] Among them, the second stage recovery cost Including gas turbine recovery costs, energy storage battery recovery costs and building microgrid compensation costs for load aggregator HVAC aggregate load regulation behavior, for The unit power compensation cost of the building microgrid to the load aggregator HVAC aggregate load at the moment, , , , In the second stage, the source load of each device The corresponding power adjustment amount at the moment; Since the second stage needs to utilize the dispatching capacity of the HVAC lumped load of the load aggregator, the second stage will no longer accurately meet the user's initial temperature setting. Therefore, formula (14) is the new constraint of the second stage relative to the first stage. are the upper and lower limits of the thermal comfort temperature of HVAC users in the building. The corresponding constraints are intended to limit the indoor temperature to the thermal comfort range of the users. Formula (15) is the decision variable for the second stage. In the second stage, the start and stop status of the gas turbine and the power purchased from the grid are no longer adjusted.

[0096] On the basis of superimposing the second-stage recourse plan on the first-stage nominal plan, the third stage aims to maximize the minimum recourse cost of the building microgrid in the worst scenario of uncertain concentration of a smaller part of the source and load, and calculates the recourse plan of the energy storage battery with faster response speed and the aggregate load of the load aggregator electric vehicle. The following equations (16)-(18) are the optimization objectives of the third stage, the newly added constraints relative to the second stage, and the decision variables, respectively.

[0097] (16)

[0098] (17)

[0099] (18)

[0100] Among them, the third stage recovery cost Including the cost of energy storage battery recovery and the cost of compensating the building microgrid for the load aggregator's electric vehicle aggregate load regulation behavior, for The unit power compensation cost of the building microgrid to the aggregate load of electric vehicles of the load aggregator at the moment, , , For the source load of each device in the third stage The corresponding power regulation amount at the moment; Since the third stage needs to utilize the dispatching capacity of the aggregated load of electric vehicles of the load aggregator, the third stage will no longer accurately meet all its target charging amounts. Therefore, formula (17) is the new constraint added in the third stage relative to the second stage, , The upper and lower limits of the aggregate target charging capacity of electric vehicles; , For the The upper and lower limits of the deviation of the target charging amount that an electric vehicle user can tolerate. Formula (18) is the decision variable of the second stage. In the third stage, the gas turbine output power and the HVAC lumped load power will no longer be adjusted.

[0101] Step 3: Establish a daily real-time optimization scheduling layer based on model reinforcement learning.

[0102] Firstly, the intraday real-time optimization scheduling problem is constructed as a Markov decision process, and the LSTM is used to learn the state transition model of the building microgrid system. Secondly, the day-ahead scheduling plan obtained in step 2 is used as a reference, and the optimal strategy for intraday real-time optimization scheduling is learned based on the actor-critic network and the model predictive control algorithm to compensate for the actual source and load prediction deviation.

[0103] Since the source-load prediction deviation decreases as the prediction time shortens, the day-ahead scheduling plan will deviate during the day, so it is necessary to adjust different source-load side devices to ensure the stable operation of the building microgrid. Due to the random characteristics of the source-load prediction deviation, the traditional mathematical programming-based method is not enough to deal with this random characteristic, resulting in the inability to effectively eliminate the operation deviation. Reinforcement learning based on the Markov decision process can effectively deal with the randomness of the environment. The present invention proposes to establish a real-time optimization scheduling within the day based on model reinforcement learning to effectively eliminate the operation deviation.

[0104] First, the intraday real-time optimization scheduling problem is constructed as a Markov decision process, and the state, action, and reward are defined. The state vector is defined as the following formula (19).

[0105] (19)

[0106] in, To optimize the scheduling time variables in real time during the day, To optimize the scheduling time scale in real time within the day, , , for The real-time forecast value of the source load at the moment. is the total initial power value of the electric vehicle; for Energy storage batteries store energy at all times. This is the day-ahead scheduling result.

[0107] Defining the action vector It is the following formula (20):

[0108] (20)

[0109] in, For the source load of each device The corresponding power adjustment amount at each moment.

[0110] The reward function is defined as follows (21):

[0111] (twenty one)

[0112] in, and They are The penalty cost per unit of grid power purchase violation and the penalty cost per unit of photovoltaic power reduction at each moment are: for The unit power compensation cost of the building microgrid to the aggregate load of electric vehicles of the load aggregator at the moment, for The unit power compensation cost of the building microgrid to the load aggregator HVAC aggregate load at the moment, for The power selling price of the power grid at that moment.

[0113] Using LSTM to learn the state transition model of building microgrid system , the input of the state transition model is the state and actions , the output is the state change .

[0114] (twenty two)

[0115] Integrate the idea of ​​model predictive control algorithm and use learning deterministic strategy The actor network interacts with the learned building microgrid system state transition model in a way of optimizing the finite prediction interval. The accumulated reward within the finite prediction interval obtained by the interaction is used as the update target of the critic network. The actor and critic network model parameters are continuously updated to finally obtain the optimal control strategy. .

[0116] (twenty three)

[0117] (twenty four)

[0118] (25)

[0119] in, is the critic network, formula (23) is the cumulative reward within the prediction interval of the H-step model prediction; formula (24) is the loss function of the critic network; formula (25) is the deterministic policy gradient of the actor network.

[0120] Finally, after obtaining the daily source load forecast value every 5 minutes, the dispatch center will use the actor network Get real-time scheduling plan within the day.

[0121] (26)

[0122] in, To optimize the scheduling plan in real time during the day, is the day-ahead scheduling plan, It is the intraday real-time optimal scheduling plan output by the actor network.

[0123] It should be noted that the state vector of reinforcement learning (Equation 19) contains the day-ahead scheduling plan of the second stage. In addition, the intraday real-time scheduling plan obtained by Equation 26 is the sum of the day-ahead scheduling plan and the results obtained by intraday real-time optimization.

[0124] In this implementation example, a more accurate uncertainty set is established based on data-driven. This implementation example establishes a source-load uncertainty set through a data-driven method, utilizes the probability information and prediction deviation of historical data, captures the conditional correlation between the prediction error and the prediction value, and enables the source-load uncertainty set to more flexibly adapt to changes in the source-load prediction results, thereby making the uncertainty set more accurate and reducing the conservatism of the traditional method of establishing the uncertainty set.

[0125] In this implementation example, the day-ahead scheduling based on three-stage robust optimization improves the robustness of the system. In view of the multi-time scale characteristics of the source-load side equipment and source-load uncertainty of the building microgrid, the source-load uncertainty set is decomposed into two parts with large fluctuation amplitude and small fluctuation amplitude. Based on this, a three-stage day-ahead robust optimization scheduling model is established. Compared with the traditional two-stage day-ahead robust optimization scheduling, the present invention utilizes the multi-time scale characteristics of the source-load side equipment and source-load uncertainty of the building microgrid, greatly reduces the conservatism and improves the robustness.

[0126] In this implementation example, the intraday real-time optimization scheduling based on model reinforcement learning eliminates deviations and improves scheduling efficiency. Since the source-load prediction deviation decreases with the shortening of the prediction time scale, the day-ahead scheduling plan will deviate when it is running intraday, so it is necessary to adjust different source-load side devices to ensure the stable operation of the building microgrid. Due to the random characteristics of the source-load prediction deviation, the traditional mathematical programming-based method is not enough to deal with this random characteristic, resulting in the inability to effectively eliminate the operation deviation. Reinforcement learning based on the Markov decision process can effectively deal with the randomness of the environment. The present invention proposes that the intraday real-time optimization scheduling based on model reinforcement learning can effectively eliminate the operation deviation, and this intelligent optimization has real-time adjustment capabilities, so that the building microgrid can more flexibly respond to various complex and dynamic operating environments.

[0127] Embodiment 2

[0128] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the program.

[0129] Embodiment 3

[0130] The purpose of this embodiment is to provide a computer-readable storage medium.

[0131] A computer-readable storage medium stores a computer program, which executes the steps of the above method when executed by a processor.

[0132] Embodiment 4

[0133] The purpose of this embodiment is to provide a building microgrid robust optimization scheduling system, including:

[0134] The source load uncertainty set construction module is configured to: generate a number of source load samples based on the day-ahead source load prediction value, representing the possible realization value of the source load on the next day, thereby constructing a source load uncertainty set;

[0135] The day-ahead scheduling layer establishment module is configured to: establish the day-ahead scheduling layer, including: dividing the source-load uncertainty set into a larger source-load uncertainty set and a smaller source-load uncertainty set, and establishing a day-ahead three-stage robust optimization scheduling model, wherein the day-ahead three-stage robust optimization scheduling model includes:

[0136] Phase 1: Determine the dispatch plan under the nominal scenario of source load forecast value;

[0137] Phase II: To maximize the minimum recourse cost of the building microgrid under the worst scenario of a larger source-load uncertainty concentration, the recourse plan of the gas turbine, energy storage battery, and the HVAC and electric vehicle aggregate loads of the load aggregator with different response speeds is calculated;

[0138] The second-stage recourse plan is superimposed on the nominal plan of the first stage. In the third stage, the recourse plan of the energy storage battery with faster response speed and the aggregate load of the electric vehicle of the load aggregator is calculated with the goal of maximizing the minimum recourse cost of the building microgrid under the worst scenario of uncertain concentration of a smaller part of the source and load;

[0139] The intraday real-time optimization scheduling module is configured as follows: on the basis of the day-ahead scheduling layer, an intraday real-time optimization scheduling layer is established based on model reinforcement learning, and the output is the intraday real-time optimal scheduling plan.

[0140] Embodiment 5

[0141] The purpose of this embodiment is to provide a computer program product containing instructions, which, when running on a computer, enables the computer to execute the methods and functions involved in any of the above embodiments.

[0142] The steps involved in the apparatus of the above embodiment correspond to the method embodiment 1, and the specific implementation method can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0143] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0144] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A robust optimization scheduling method for building microgrids, characterized in that: include: Based on the day-ahead source load forecast value, several source load samples are generated to represent the possible realized value of the source load on the next day, thereby constructing a source load uncertainty set; Establishing a day-ahead scheduling layer includes: dividing the source-load uncertainty set into a larger source-load uncertainty set and a smaller source-load uncertainty set, and establishing a day-ahead three-stage robust optimization scheduling model, wherein the day-ahead three-stage robust optimization scheduling model includes: Phase 1: Determine the dispatch plan under the nominal scenario of source load forecast value; Phase II: To maximize the minimum recourse cost of the building microgrid under the worst scenario of a larger source-load uncertainty concentration, the recourse plan of the gas turbine, energy storage battery, and the HVAC and electric vehicle aggregate loads of the load aggregator with different response speeds is calculated; The second-stage recourse plan is superimposed on the nominal plan of the first stage. In the third stage, the recourse plan of the energy storage battery with faster response speed and the electric vehicle aggregate load of the load aggregator is calculated with the goal of maximizing the minimum recourse cost of the building microgrid in the worst scenario of uncertain concentration of a smaller part of the source and load. Finally, the dispatch plans of the three stages are summarized as the day-ahead dispatch plan and sent to the lower-level intraday real-time optimization dispatch layer. On the basis of the day-ahead scheduling layer, the intraday real-time optimization scheduling layer is established based on model reinforcement learning to output the intraday real-time optimal scheduling plan; The specific process of constructing the source-load uncertainty set is as follows: Get the latest day-ahead forecast value; Obtain the expected value and covariance matrix of the conditional correlation between the source load prediction value and the true value; The conditional correlation between the source load prediction value and the true value is sampled to generate a number of samples, and then a sample of the source load true value is obtained; based on The samples obtained by sampling at the moment and the latest day-ahead forecast value establish a source load uncertainty set; Among them, the first stage: determine the dispatch plan under the nominal scenario of source load forecast value, including the first stage optimization objectives, constraints and decision variables; The constraints in the first stage include power balance constraints, gas turbine operation constraints, grid constraints, energy storage battery operation constraints, and load aggregator constraints; The optimization target of the second stage is the second stage recovery cost, which specifically includes the gas turbine recovery cost, the energy storage battery recovery cost and the compensation cost of the building microgrid for the load aggregator's HVAC aggregate load regulation behavior; The new constraints added in the second stage compared to the first stage are the upper and lower limits of thermal comfort temperature for HVAC users in the building. The corresponding constraints are intended to limit the indoor temperature to the thermal comfort range of users; The third stage recourse costs include the energy storage battery recourse costs and the compensation costs of the building microgrid for the load aggregator's electric vehicle aggregate load regulation behavior; Compared with the second stage, the third stage adds the upper and lower limits of the aggregate target charging capacity of electric vehicles; The upper and lower limits of the deviation from the target charging amount that an electric vehicle user can tolerate; Based on model reinforcement learning, a real-time optimization scheduling layer is established within the day, including: The intraday real-time optimization scheduling problem is constructed as a Markov decision process, defining states, actions, and rewards; The state transition model of the building microgrid system is learned using LSTM. The input of the state transition model is the state and action, and the output is the state change. The actor network that learns deterministic strategies interacts with the learned building microgrid system state transition model in a way that optimizes the finite prediction interval. The accumulated reward within the finite prediction interval obtained by the interaction is used as the update target of the critic network. The actor and critic network model parameters are continuously updated to finally obtain the optimal control strategy. After obtaining the daily source-load forecast value at set intervals, the dispatch center will use the actor network optimal control strategy to obtain the daily real-time dispatch plan.

2. A method for robust optimization scheduling of a building microgrid as claimed in claim 1, characterized in that: The source load uncertainty set is decomposed into an uncertainty set with large fluctuation amplitude and long time scale characteristics and an uncertainty set with small fluctuation amplitude and short time scale characteristics.

3. A building microgrid robust optimization dispatching system, characterized in that: include: The source load uncertainty set construction module is configured to: generate a number of source load samples based on the day-ahead source load prediction value, representing the possible realization value of the source load on the next day, thereby constructing a source load uncertainty set; The day-ahead scheduling layer establishment module is configured to: establish the day-ahead scheduling layer, including: dividing the source-load uncertainty set into a larger source-load uncertainty set and a smaller source-load uncertainty set, and establishing a day-ahead three-stage robust optimization scheduling model, wherein the day-ahead three-stage robust optimization scheduling model includes: Phase 1: Determine the dispatch plan under the nominal scenario of source load forecast value; Phase II: To maximize the minimum recourse cost of the building microgrid under the worst scenario of a larger source-load uncertainty concentration, the recourse plan of the gas turbine, energy storage battery, and the HVAC and electric vehicle aggregate loads of the load aggregator with different response speeds is calculated; The second-stage recourse plan is superimposed on the nominal plan of the first stage. In the third stage, the recourse plan of the energy storage battery with faster response speed and the electric vehicle aggregate load of the load aggregator is calculated with the goal of maximizing the minimum recourse cost of the building microgrid in the worst scenario of uncertain concentration of a smaller part of the source and load. Finally, the dispatch plans of the three stages are summarized as the day-ahead dispatch plan and sent to the lower-level intraday real-time optimization dispatch layer. The intraday real-time optimization scheduling module is configured as follows: on the basis of the day-ahead scheduling layer, the intraday real-time optimization scheduling layer is established based on model reinforcement learning, and the intraday real-time optimal scheduling plan is output; The specific process of constructing the source-load uncertainty set is as follows: Get the latest day-ahead forecast value; Obtain the expected value and covariance matrix of the conditional correlation between the source load prediction value and the true value; The conditional correlation between the source load prediction value and the true value is sampled to generate a number of samples, and then a sample of the source load true value is obtained; based on The samples obtained by sampling at the moment and the latest day-ahead forecast value establish a source load uncertainty set; Among them, the first stage: determine the dispatch plan under the nominal scenario of source load forecast value, including the first stage optimization objectives, constraints and decision variables; The constraints in the first stage include power balance constraints, gas turbine operation constraints, grid constraints, energy storage battery operation constraints, and load aggregator constraints; The optimization target of the second stage is the second stage recovery cost, which specifically includes the gas turbine recovery cost, the energy storage battery recovery cost and the compensation cost of the building microgrid for the load aggregator's HVAC aggregate load regulation behavior; The new constraints added in the second stage compared to the first stage are the upper and lower limits of thermal comfort temperature for HVAC users in the building. The corresponding constraints are intended to limit the indoor temperature to the thermal comfort range of users; The third stage recourse costs include the energy storage battery recourse costs and the compensation costs of the building microgrid for the load aggregator's electric vehicle aggregate load regulation behavior; Compared with the second stage, the third stage adds the upper and lower limits of the aggregate target charging capacity of electric vehicles; The upper and lower limits of the deviation from the target charging amount that an electric vehicle user can tolerate; Based on model reinforcement learning, a real-time optimization scheduling layer is established within the day, including: The intraday real-time optimization scheduling problem is constructed as a Markov decision process, defining states, actions, and rewards; The state transition model of the building microgrid system is learned using LSTM. The input of the state transition model is the state and action, and the output is the state change. The actor network that learns deterministic strategies interacts with the learned building microgrid system state transition model in a way that optimizes the finite prediction interval. The accumulated reward within the finite prediction interval obtained by the interaction is used as the update target of the critic network. The actor and critic network model parameters are continuously updated to finally obtain the optimal control strategy. After obtaining the daily source-load forecast value at set intervals, the dispatch center will use the actor network optimal control strategy to obtain the daily real-time dispatch plan.

4. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method described in any one of claims 1 to 2 are implemented.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 2 are performed.

Images