Green power direct connection energy supply optimization method and system considering uncertainty and system toughness

Abstract

The invention relates to the technical field of energy supply scheme optimization, and provides a green power direct connection energy supply optimization method and system considering uncertainty and system toughness, and the method comprises the steps: constructing a Transform model based on time fusion, obtaining the operation data of a green power direct connection power supply system, carrying out the load prediction and upstream green power supply prediction, and calculating the uncertainty; a multi-objective optimization model is constructed by taking the minimization of the total operation cost, the minimization of the equipment power fluctuation and the maximization of the system energy supply toughness as optimization objectives; aiming at the multi-objective optimization model, solving a Pareto optimal scheduling scheme set by adopting a crowd-dwelling adaptive collaborative algorithm so as to regulate and control the operation scheduling of the green power direct connection power supply system; an IES-CACA algorithm of a multi-sub-population coevolution and adaptive evolution strategy is introduced, and an uncertainty perception regulation and control mechanism is combined, so that the global optimization capability and the system toughness of multi-target optimization scheduling in a green power direct connection scene are improved.

Description

Technical Field

[0001] This invention relates to the technical field of energy supply scheme optimization, specifically to a method and system for optimizing green electricity direct connection energy supply that takes into account uncertainty and system resilience. Background Technology

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

[0003] As a significant contributor to energy consumption and carbon emissions, the building sector is facing an urgent need to transform towards green, low-carbon, efficient, and intelligent systems. Office buildings, due to their large energy loads, significant peak-valley variations, and high operational stability requirements, have become a key scenario for optimizing integrated energy systems. In recent years, the "green direct connection" model has gained widespread attention in the building sector. This model directly couples office buildings to nearby distributed wind or photovoltaic power systems, enabling the local consumption of clean energy and avoiding the intermediate links of the traditional power grid. This model helps reduce transmission and distribution losses, improves energy utilization efficiency, and provides an effective path to achieve near-zero carbon operation of office buildings. However, under the green direct connection model, the uncertainty on both sides of building load and renewable energy output increases dramatically, and system scheduling faces more complex constraints and conflicting objectives. Therefore, an optimization method that balances multi-objective performance, scheduling stability, and operational resilience is urgently needed to achieve comprehensive synergy in the green, economic, and reliable dimensions of office building energy systems.

[0004] However, existing optimization methods still face multiple technical bottlenecks when dealing with high-dimensional nonlinear modeling and dynamic operation scheduling problems in the context of "green electricity direct connection" in office buildings: First, there is a trade-off between model solution efficiency and accuracy. The optimization problem of integrated energy systems in office buildings is typically characterized by high dimensionality, strong nonlinearity, and multiple constraints. Traditional deterministic optimization algorithms are prone to getting trapped in local optima and struggle to effectively search within complex solution spaces. While general metaheuristic algorithms possess global optimization capabilities, they often struggle to balance convergence speed with solution quality, failing to meet the dual requirements of real-time performance and executability for building energy system scheduling. Second, existing optimization strategies struggle to effectively coordinate solution set diversity and convergence. Traditional archiving and updating mechanisms primarily rely on distance or density indices, making it difficult to maintain a balanced distribution of Pareto solutions in the solution space. This results in limited coverage and insufficient flexibility of the generated optimization schemes in practical applications. Furthermore, the collaborative optimization capability among multiple objectives is significantly insufficient. In green electricity direct connection scenarios, economic efficiency, operational stability, and system resilience often conflict (e.g., high resilience may lead to increased operating costs). Traditional optimization models often prioritize single objectives or simple weighted summaries, making it difficult to form a comprehensively balanced Pareto front solution set, thus limiting the strategy space for dynamic scheduling. More seriously, most existing algorithms lack modeling and adaptation mechanisms for source load uncertainties, and fail to fully consider the impact of wind and solar power output fluctuations and building load uncertainties on scheduling results. This results in insufficient robustness and poor stability of the optimization scheme in actual implementation, seriously affecting the energy supply security of critical loads. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a green electricity direct connection energy supply optimization method and system that considers uncertainty and system resilience. It introduces the IES-CACA algorithm, which employs multi-subpopulation collaborative evolution and adaptive evolution strategies, and combines it with an uncertainty-aware control mechanism to enhance the global optimization capability and system resilience of multi-objective optimization scheduling in green electricity direct connection scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: One or more embodiments provide an optimized method for green electricity direct connection energy supply that takes into account uncertainties and system resilience, including the following steps: A time-fusion Transformer model is constructed to obtain the operating data of the green electricity direct-connection power supply system for load forecasting and upstream green electricity supply forecasting, and to calculate the uncertainty. A multi-objective optimization model is constructed with the optimization objectives of minimizing total operating cost, minimizing equipment power fluctuations, and maximizing system energy supply resilience. For the multi-objective optimization model, the swarm adaptive cooperative algorithm is used to solve the Pareto optimal scheduling scheme set to regulate the operation and scheduling of the green electricity direct-connection power supply system. In the process of solving the problem using the gregarious adaptive cooperative algorithm, during the initialization phase, a corresponding number of subpopulations and an external archive are constructed according to the number of objectives in the multi-objective optimization model. Each subpopulation is used to optimize one objective. An initial subpopulation is generated based on a hybrid initialization strategy that matches historical scenarios. During the population evolution phase, an adaptive evolution strategy is adopted, and evolutionary opportunities are dynamically allocated based on the optimization contribution of each subpopulation and the uncertainty information of the prediction.

[0007] One or more embodiments provide a green direct-connection energy supply optimization system that takes into account uncertainty and system resilience, including: The prediction module is configured to build a time-fusion Transformer model, obtain the operating data of the green electricity direct-connection power supply system to perform load prediction and upstream green electricity supply prediction, and calculate uncertainty; The model building module is configured to construct a multi-objective optimization model with the optimization objectives of minimizing total operating cost, minimizing equipment power fluctuations, and maximizing system energy supply resilience. The solution module is configured to use a swarm adaptive cooperative algorithm to solve for the Pareto optimal scheduling scheme set for a multi-objective optimization model, so as to regulate the operation and scheduling of the green electricity direct-connection power supply system. In the process of solving the problem using the gregarious adaptive cooperative algorithm, during the initialization phase, a corresponding number of subpopulations and an external archive are constructed according to the number of objectives in the multi-objective optimization model. Each subpopulation is used to optimize one objective. An initial subpopulation is generated based on a hybrid initialization strategy that matches historical scenarios. During the population evolution phase, an adaptive evolution strategy is adopted, and evolutionary opportunities are dynamically allocated based on the optimization contribution of each subpopulation and the uncertainty information of the prediction.

[0008] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the above-described method for optimizing green direct-connection power supply considering uncertainties and system resilience.

[0009] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps in the above-described method for optimizing green direct-connection power supply considering uncertainties and system resilience.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention significantly enhances the global optimization capability of integrated energy systems in office buildings under high-dimensional, highly nonlinear, and complex multi-constraint scenarios by introducing the IES-CACA optimization algorithm. IES-CACA utilizes a multi-subpopulation co-evolutionary mechanism, ensuring each subpopulation focuses on a single optimization objective. This guarantees multi-objective collaborative balance while improving the diversity and distribution equilibrium of the solution set. Simultaneously, the adaptive evolutionary strategy dynamically allocates evolutionary opportunities based on the optimization contribution of each subpopulation in the current iteration and the uncertainty information predicted by the Transformer model, effectively overcoming the trade-off between convergence speed and accuracy inherent in traditional metaheuristic algorithms. Compared to traditional methods, this invention significantly shortens optimization computation time while maintaining system scheduling accuracy, improving the real-time performance, robustness, and executability of the multi-objective optimization model in dynamic environments. This method is particularly suitable for solving scheduling optimization problems with significant load and energy supply fluctuations and significant objective conflicts in green power direct connection scenarios, enhancing the resilience of green power direct connection power supply systems in the face of uncertainties in wind and solar power output and the stability requirements of critical loads.

[0011] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0013] Figure 1 This is a flowchart of the energy supply optimization method according to Embodiment 1 of the present invention; Figure 2 This is a flowchart illustrating the solution process of the gregarious adaptive cooperative algorithm in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of the green electricity direct connection energy supply optimization system considering uncertainty and system resilience in Embodiment 3 of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0017] Example 1 In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 3 As shown, the optimization method for green electricity direct connection energy supply considering uncertainties and system resilience includes the following steps: Step 1: Construct a time-fusion Transformer model to obtain point predictions and uncertainty quantification results for future load and green power output.

[0018] Based on the time fusion Transformer model, quantile regression can be used for training, enabling it to output different quantiles of the predicted target, thereby characterizing the possible distribution range of the predicted value.

[0019] Step 2: Construct a multi-objective optimization model with the optimization objectives of minimizing total operating cost, minimizing equipment power fluctuations, and maximizing system power supply resilience; Step 3: For the multi-objective optimization model, the swarm adaptive cooperative algorithm is used to solve for the Pareto optimal scheduling scheme set in order to regulate the operation and scheduling of the green electricity direct power supply system. In the process of solving the problem using the gregarious adaptive cooperative algorithm, during the initialization phase, a corresponding number of subpopulations and an external archive are constructed according to the number of objectives in the multi-objective optimization model. Each subpopulation is used to optimize one objective. An initial subpopulation is generated based on a hybrid initialization strategy that matches historical scenarios. During the population evolution phase, an adaptive evolution strategy is adopted, and evolutionary opportunities are dynamically allocated based on the optimization contribution of each subpopulation and the uncertainty information of the prediction.

[0020] This implementation method introduces a time-fusion Transformer prediction model to achieve high-precision prediction of load and green electricity supply in a green electricity direct-connection power supply system for office buildings, thereby estimating the uncertainty in each time period. The Transformer model has the ability to capture long-term dependencies and complex temporal relationships, and is suitable for energy data with high volatility and obvious seasonal characteristics. In the optimization modeling stage, a multi-objective scheduling model with a triple objective function is established: minimizing operating costs, stabilizing power output, and maximizing system energy supply resilience are the optimization objectives. Based on this model, a gregarious adaptive cooperative algorithm is introduced for solving the problem. The algorithm adopts a multi-subpopulation structure, with each subpopulation independently optimizing one objective, improving the diversity of solutions and the level of cooperation between objectives. During initialization, historical operating data and scenario patterns are combined to generate an initial population with strong coverage through a hybrid approach. During the evolution process, the evolutionary resources of each subpopulation are dynamically adjusted according to its contribution to the overall Pareto front and the current level of uncertainty, thereby achieving adaptive regulation and flexible switching of search strategies. Finally, a Pareto solution set with high fitness is generated to guide the dynamic scheduling and energy allocation of the green electricity direct-connection power supply system.

[0021] This implementation significantly enhances the global optimization capability of integrated energy systems in office buildings under high-dimensional, highly nonlinear, and complex multi-constraint scenarios by introducing the IES-CACA optimization algorithm. IES-CACA utilizes a multi-subpopulation co-evolutionary mechanism, ensuring each subpopulation focuses on a single optimization objective. This guarantees multi-objective collaborative balance while improving the diversity and distribution equilibrium of the solution set. Simultaneously, the adaptive evolutionary strategy dynamically allocates evolutionary opportunities based on the optimization contribution of each subpopulation in the current iteration and the uncertainty information predicted by the Transformer model, effectively overcoming the trade-off between convergence speed and accuracy inherent in traditional metaheuristic algorithms. Compared to traditional methods, this invention significantly shortens optimization computation time while maintaining system scheduling accuracy, improving the real-time performance, robustness, and executability of the multi-objective optimization model in dynamic environments. This method is particularly suitable for solving scheduling optimization problems with significant load and energy supply fluctuations and significant objective conflicts in green power direct connection scenarios, enhancing the resilience of green power direct connection power supply systems in the face of uncertain wind and solar power output and critical load stability requirements.

[0022] Step 1: Green electricity direct connection power supply system, including power plants that provide green electricity and integrated energy system for regulating and distributing power supply. The integrated energy system connects the power grid and the upstream power plants that provide green electricity, and also connects to heat energy, cold energy and gas sources, etc., and performs energy conversion and dispatching through the integrated energy system; The data acquisition in step 1 provides the data foundation for the entire optimization method. The operational data of the green electricity direct-connection power supply system includes: meteorological and environmental data, upstream green electricity supply basic data, operating parameters of key equipment in the integrated energy system, energy consumption data of the target area, and market and resilience parameters. Meteorological and environmental data, including historical and predicted ambient temperature, wind speed, and solar radiation intensity of the target area and the upstream Fengguang Industrial Park; Upstream green electricity supply basic data includes wind and solar power installed capacity, equipment power curves, historical hourly power generation in wind and solar industrial parks, and negotiated electricity price information through direct connection channels; Operating parameters of key equipment in the target area, such as office buildings, including the rated power and heating / cooling coefficient of performance curves of medium-deep ground source heat pumps, the rated cooling / heating power and energy efficiency ratio of electric air conditioners, and the rated capacity, maximum charging and discharging power, charging and discharging efficiency and self-discharge rate of electrochemical energy storage systems.

[0023] Energy consumption data and related data for the target area, such as operational data for office buildings, including historical hourly electricity, cooling, and heating load data, as well as date-type characteristics reflecting the patterns of human activity.

[0024] Market and resilience parameters include punitive tariffs on grid backup power and critical load levels and minimum guaranteed power durations required for system resilience design.

[0025] In step 1, a time-fusion-based Transformer model is constructed, including a data preprocessing and covariate partitioning module, multiple GRN networks, a sequence encoder, and a decoder; The time-fusion Transformer model constructed in this embodiment is a hybrid prediction model. The TFT model processes static metadata, known future information and historical observation data through its unique architecture to quantify the uncertainty of load and green power output in the short term. Step 1 involves constructing a time-fusion Transformer model for load forecasting and upstream green electricity supply forecasting, and calculating uncertainties. This method includes the following steps: Step 11, Data Preprocessing and Covariate Partitioning: Clean and normalize the data, and then partition the processed data into covariates; The acquired raw historical time-series data is cleaned, outliers are removed, and missing values ​​are linearly interpolated; for any feature k, its raw observation at time step t is denoted as... To eliminate the influence of dimensions and accelerate model convergence, a max-min normalization method is used to scale each feature to the [0,1] interval. The normalized feature values ​​are... The calculation is as follows: ; in, T is the set of time steps in the training set. Subsequently, according to the input requirements of the TFT model, the normalized features are strictly divided into three types of covariate inputs; covariate partitioning involves classifying the normalized data according to variable type to obtain the partitioned covariates, including static covariates, known future covariates, and historical observation covariates; static covariates Features that remain constant within each predicted sample (i.e., a predicted sequence), such as building type encoding, have a dimension of 1. ; Known future covariates When predicting any future time t, the features whose value is known include, for example, the hour of time t, the day of the week, and holiday markers. The dimension of these features is... This vector will be used as input to the GRN network and transformed into feature representations. .

[0026] Historical observation covariates The features whose future values ​​are to be predicted are only known in the past, such as historical load values, historical wind and solar power output, and historical temperature. The dimensions are: This vector will be used as input to the GRN network and transformed into feature representations. .

[0027] Step 12, Variable Selection and Feature Encoding: For static covariates, a static encoder composed of multiple GRNs is used to encode the static covariates into multiple context vectors; static covariates Context vectors are generated using a static encoder. For example, context vectors used for variable selection. We obtain it from the following formula:

[0028] For both historical observed covariates and known future covariates, GRN transformation, feature concatenation, weight calculation, and weighted summation are performed sequentially to obtain the final historical feature representation. With known future characteristics : Step 121, GRN Transformation: Transform each type of covariate using a gated recurrent network to obtain the original feature representation of the covariate; Specifically, for each type of covariate, it is first transformed through an independent gated recurrent network (GRN); Taking historical observation covariates as an example, the corresponding original feature representations are obtained through GRN transformation. The GRN transformation process is as follows: ; in, It is a gated residual network specifically designed for historical covariates.

[0029] Similarly, given future covariates The original feature representation is obtained through GRN transformation.

[0030] Step 122: Calculate the weights based on the concatenated vector after covariate transformation and the context vector for feature selection encoded by static covariates. Then, weight the original feature representations of the covariates based on the weights to obtain the final feature representations. First, the original characteristics of the historical observed covariates are characterized. Characterization of original features of known future covariates Concatenate them into a joint feature vector : ; Combine the joint feature vector with the context vector obtained from static covariate encoding for feature selection. The weight vector is selected by calculating variables through another GRN. :

[0031] Among them, the weight vector Dimensions and joint feature vectors The same, each of its elements Corresponding to The Middle k The weights of each feature.

[0032] Weight vector Break it down into covariates corresponding to historical observations. and corresponding to the known future covariates Assuming The dimension is , The dimension is ,but:

[0033] Ultimately, the weighted historical characteristics are represented. Characterized by the original features of historical observation covariates Its corresponding weight The sum of the elements is obtained by multiplying each element and then summing them: ; in, Representing vectors The k The element (i.e., the element)k A historical characteristic.

[0034] Similarly, weighted representation of known future features Characterized by the original features of known future covariates Its corresponding weight The sum of the elements is obtained by multiplying each element and then summing them: ; in, Representing vectors The k The element (i.e., the element) k (One known feature in the future).

[0035] The above steps utilize a variable selection network to screen the most relevant input features for each type of covariate, thereby improving model resilience. This embodiment uses a gated recurrent network (GRN) to achieve feature transformation and weight selection, combined with the contextual information of static covariates, to improve the adaptability and resilience of the time-fusion-based Transformer model to different types of features.

[0036] Step 13, Temporal Feature Extraction: Representing the processed historical features The input is fed into a sequence encoder for temporal feature extraction to capture local temporal patterns and obtain encoded features; The initial state of this sequence encoder can be set by a static context vector, allowing its output to adapt to a specific scenario. The encoder output contains a hidden state sequence {ht} containing past information.

[0037] Step 14: For the encoded features, perform an interpretable multi-head attention operation, perform a one-dimensional linear transformation and temporal fusion, and combine it with known future features. By fusing the data, the hidden states at each future time step can be decoded. This is the core of the TFT model for capturing long-term dependencies. First, the hidden state sequence {ht} output by the encoder is used as the key K and value V, and the decoder's query vector for the future is used as the query Q. An interpretable multi-head attention mechanism is then executed as follows: For the h-th attention head, its attention weight is calculated as follows: ; in, , , Let be the learnable parameter matrix of the h-th head, and dk be the dimension of the key vector. In interpretable multi-head attention, the outputs of multiple heads are averaged and then a linear transformation is applied to obtain the final output: ; in, It's about the number of heads to focus on. This is the output transformation matrix.

[0038] The multi-head attention mechanism enables the model to output a unified and interpretable attention weight matrix, intuitively showing which historical moments are most important for the current prediction. The output of the attention layer is then compared with the processed known future features. The data are fused and passed through a feedforward network to finally generate the hidden state of each time step within the future prediction window, i.e., the decoded features.

[0039] Step 15, Prediction Output: The hidden state of each future time step obtained in Step 14 is mapped to the final prediction value through a fully connected output layer, thereby obtaining the quantile results of the predicted load and the predicted green power output for each time period. In this embodiment, the predicted output values ​​for the predicted load and the predicted green electricity output are each defined as an interval. Within this interval, each predicted value has a probability of being accurately predicted. This embodiment uses quantiles to represent the probability, taking a point from the interval of predicted values ​​as the 10th quantile. 50th percentile 90th percentile ; represent the probabilities that the actual value is less than the predicted value at the corresponding point, which are 10%, 50%, and 90%, respectively. For example, if the predicted output range is [A1, A2, A3], then A1 can be the 10th percentile. The probability that the actual value is less than A1 is 10%; A2 is the 50th percentile. The probability that the actual value is less than A2 is 50%; A3 is the 90th percentile. The probability that the actual value is less than A3 is 90%.

[0040] Training of the Transformer model based on time fusion: For multivariate prediction, the output layer will generate multiple channels. The Transformer model training aims to minimize the gap between the predicted and the true values. The model's ability to output multiple target quantiles stems from the multivariate quantile loss function used during the training phase. This function enables the model to simultaneously learn the conditional quantile distributions of five prediction targets (cooling, heating, and electrical loads; photovoltaic and wind power outputs).

[0041] Assume there is a total There are 1 prediction target, and its index set is 1 For each objective k∈K The Transformer model needs to learn a set of selected quantile proportions. The predicted value is below.

[0042] For a batch of training samples, its overall loss function Defined as the sum of quantile losses for all targets and all quantiles: ; in, The quantile loss function (also known as bouncing loss) is defined as follows: ; In the formula, N The number of samples in the training batch; For the first i In the nth sample, the nth k The true value of the predicted target. For the model to the first i The sample, the first k Each target output Quantile predictions. The target quantile has a value range of (0,1). For indicator functions, when It is 1 if it is true, otherwise it is 0.

[0043] Minimize the total loss using the backpropagation algorithm. The model parameters are optimized so that it learns to capture information from input features (historical data, covariates) and accurately estimate the quantiles of the five target variables at different confidence levels in the future.

[0044] Probabilistic interpretation of quantiles: For any prediction target (e.g., electrical load), the quantile prediction values ​​of its output have clear statistical significance: 10th percentile This indicates that within the predicted value range of the model output, the actual value of the k-th target in the future time period t is lower than this predicted value (10th quantile). The probability of () is 10%.

[0045] 50th percentile This indicates that, within the predicted value range of the model output, the actual value of the k-th target in the future time period t is lower than this predicted value (50th quantile). The probability of ( ) is 50%, and there is a 50% probability that it is higher than this value; 90th percentile This indicates that within the predicted value range of the model output, the actual value of the k-th target in the future time period t is lower than this predicted value (90th quantile). The probability of () is 90%.

[0046] Therefore, the interval This constitutes an 80% confidence interval for the future actual value, which intuitively represents the range of uncertainty in the prediction: the wider the interval, the higher the uncertainty.

[0047] To drive the subsequent deterministic optimization model operation, the median predicted value of each objective is used as the baseline predicted value, i.e.: Forecast load values: ; Predicted green power output: ; In the formula, Generally refers to the baseline forecast value and the median value of each forecast load at time t; Generally refers to the baseline predicted output value and the corresponding predicted median value of photovoltaic or wind power generation at time t.

[0048] Step 16: Based on the quantile results of different prediction targets for each time period, calculate the uncertainty index. ; To provide a concise and effective risk perception signal to the optimization algorithm, five independent uncertainty indicators are integrated into a single comprehensive uncertainty indicator. .

[0049] Calculate the independent uncertainty index for each objective: First, calculate the relative uncertainty of each prediction target in time period t: Cooling load: ; Heat load: ; Electrical load: ; Photovoltaic output: ; Wind power output: ; in, These are the 10th, 50th, and 90th percentile values ​​for the predicted cooling load, respectively; similarly... These are the predicted quantile values ​​when the quantiles for heat load, electrical load, photovoltaic output, and wind power output are all n%. For extremely small positive numbers, to prevent division by zero; each Each value represents the width of the 80% confidence interval for the corresponding prediction relative to its baseline value; a larger value indicates higher uncertainty.

[0050] Constructing a comprehensive uncertainty indicator: Defining the time period t Comprehensive uncertainty index The weighted average of the five independent indicators mentioned above: ; in, Let be the weight coefficients for each objective, satisfying... .

[0051] Determining the weights: The weighting coefficients reflect the degree of influence of different prediction objectives on the overall stability of system scheduling. A recommended calculation method is to normalize based on the average baseline power ratio of each objective within a typical scheduling period: ; in This represents the average value of the baseline cooling load forecast sequence over the scheduling period T; the average values ​​of other targets are calculated in the same way. Other weights And so on, the calculation method is the same.

[0052] In the above implementation, the calculation of the uncertainty index can ensure that the uncertainty of the target with high power and significant impact on system balance is dominant in the comprehensive index.

[0053] Furthermore, in order to construct a multi-objective optimization model, the integrated energy system is first modeled, and the physical system, which includes energy conversion, energy storage and system operation constraints, is transformed into a computable mathematical model, including the mathematical model of the equipment in the integrated energy system and the constraint model of the operation of the integrated energy system. Optionally, the mathematical models of equipment in the integrated energy system include medium-deep ground source heat pump models, electric air conditioning models, and electrochemical energy storage system models; As the basic heat source and cold source of a system, the coefficient of performance (COP) of a medium-deep ground source heat pump is significantly affected by the ground source temperature and the supply water temperature on the load side. The model of a medium-deep ground source heat pump needs to accurately characterize its variable operating conditions and operational constraints. Establish the coefficient of performance for heating mode: ; Among them, COP h (t) represents the heating performance coefficient for time period t; T soil (t) represents the average soil temperature (°C) at the source side during time period t; T sup,h (t) represents the design water supply temperature (°C) of the heating system during time period t. These are the fitting coefficients for the equipment characteristics.

[0054] Establish the coefficient of performance for cooling mode: ; Among them, COP c (t) represents the coefficient of performance for cooling during time period t; T sup,c (t) represents the design water supply temperature (°C) of the cooling system during time period t. These are the fitting coefficients for the equipment characteristics.

[0055] Establish the relationship between the output power and electrical power of a medium-deep ground source heat pump: ; Among them, P hp,heat (t), P hp,cool (t) represent the thermal power and cooling power (kW) output by the ground source heat pump during time period t; P hp,elec (t) represents the electrical power (kW) consumed by the ground source heat pump during time period t. Operational constraints: ; in, ∈{0,1} represents the operating status flag of the ground source heat pump during time period t (1 indicates operation); Minimum technical output and rated electrical power (kW); and These are the maximum and minimum power change rates (kW / Δt), respectively. Auxiliary variables representing the start-up and shutdown times during time period t; , Minimum continuous operation and downtime (or number of time periods).

[0056] Electric air conditioner model: As a flexible peak-shaving device, the electric air conditioner model needs to consider its rapid response characteristics and energy efficiency changes with ambient temperature, including variable operating condition energy efficiency model, the relationship between the output power and electrical power of the electric air conditioner, and operating constraints. Specifically, a variable operating condition energy efficiency model is constructed for electric air conditioners to describe the energy efficiency ratio of electric air conditioners as a function of outdoor temperature. The changes are represented by piecewise linear functions: Cooling mode: ; Heating mode: ; in, and These represent the cooling and heating energy efficiency ratios for time period t, respectively. and These are the temperature thresholds (°C) for cooling and heating segments, respectively. The fitting coefficients are denoted as .

[0057] The relationship between the output power and electrical power of an electric air conditioner: Cooling mode: ; Heating mode: ; in, and These are the cooling power and heating power (kW) output by the electric air conditioner during time period t, respectively. The electrical power consumed during time period t (kW).

[0058] Operational constraints: ; in, This is a status indicator for the electric air conditioner during time period t. Minimum technical output and rated electrical power (kW); T represents the absolute value of the maximum permissible power change (kW / Δt). ac,min Minimum continuous running time (number of time slots); (.) + This indicates that a positive value is taken.

[0059] Specifically, the electrochemical energy storage system model includes the electrochemical energy storage equation of state and operating constraints; The electrochemical energy storage equation of state describes the charge state update process, and its mathematical model is expressed as follows: ; Operating power and operating constraints include the following: ; Wherein, SOC(t) represents the state of charge of the electrochemical energy storage system during time period t; Self-discharge rate; For charging and discharging efficiency; The charging and discharging power (kW) during time period t; Rated capacity (kWh); Maximum charging and discharging power (kW); This serves as a charge / discharge status indicator. These are the upper and lower limits of the state of charge, respectively.

[0060] In some embodiments, the constraint model for the operation of the integrated energy system includes energy balance constraints, equipment operation constraints, and grid interaction constraints; Optionally, energy balance constraints include electrical balance constraints, thermal balance constraints, and cooling balance constraints. Power balance constraints: ; Among them, P grid (t) represents the power (kW) purchased through the green electricity direct connection channel during time period t; P PV (t) represents the photovoltaic output during time period t, such as the photovoltaic output (kW) of an office building rooftop; P load,e (t) represents the electrical load (kW) of the office building during time period t.

[0061] Thermodynamic equilibrium constraints: ; Among them, P load,h (t) represents the heat load (kW) of the target area during time period t; Cooling balance constraints: ; Among them, P load,c (t) represents the cooling load (kW) of the target area, such as office buildings, during time period t.

[0062] Optional equipment operating constraints include: Equipment output upper and lower limit constraints: ; in, These are the minimum technical output and rated power (kW) of equipment d, respectively. Let d be the output of device d at time t.

[0063] Equipment ramping constraints: ; in, These represent the maximum allowable power increase and decrease rates (kW / Δt) within a single scheduling period for device d.

[0064] Optional, grid interaction constraints include: Direct power purchase power constraints: ; in, These are the wind power and photovoltaic power (kW) that can be supplied by the upstream wind and solar park during time period t, as predicted in step 1. The maximum capacity (kW) for direct green electricity connection channels; This is for directly connected power purchase; this constraint ensures that the purchased power does not exceed the available green electricity and transmission capacity of the upstream lines.

[0065] In step 2, a multi-objective optimization model is constructed to achieve synergistic optimization of economy (F1), stability (F2) and resilience (F3), including minimizing the total operating cost (F1), minimizing the power fluctuation of key equipment (F2), and maximizing the system power supply resilience (F3). Specifically, minimizing the total operating cost (F1) and the overall operating cost of the system within a scheduling cycle will reduce the cost of purchasing green electricity. Equipment operation and maintenance costs Equipment depreciation costs Minimizing the total cost is expressed as follows: ; In the formula: T is the total number of time periods in one scheduling cycle; π grid(t) represents the electricity purchase price (yuan / kWh) during time period t via the green electricity direct connection channel; P grid (t) represents the purchased power (kW) during time period t, which is a decision variable; Δt represents the duration (hours) of a single dispatch period. This includes all operating equipment, including medium-deep ground source heat pumps (HP), electric air conditioning (AC), and electrochemical energy storage systems (BES). P represents the unit power operation and maintenance cost coefficient (yuan / kWh) for equipment d; d (t) represents the operating power (kW) of device d during time period t. For electrochemical energy storage systems, it represents the sum of the absolute values ​​of charging and discharging power. This refers to a set of critical equipment that is sensitive to power fluctuations, typically including ground source heat pumps and electrochemical energy storage systems; This represents the power fluctuation depreciation cost coefficient for device d; Let d be the rated power (kW) of the device.

[0066] Specifically, the power fluctuation term (F2) of critical equipment is minimized, which minimizes the rate of power change of critical equipment between adjacent scheduling periods to quantify operational stability and extend equipment life. The formula is expressed as follows: ; In the formula: The power consumed by the ground source heat pump during time period t (kW). The net charge and discharge power of the electrochemical energy storage system during time period t. (kW); These are the rated electrical power (kW) of the ground source heat pump and the electrochemical energy storage system, respectively.

[0067] Specifically, maximize the system energy resilience (F3), which maximizes the shortest time that the system can maintain the operation of critical loads in office buildings by relying solely on energy storage in the event of a green electricity supply interruption.

[0068] To facilitate optimization, it is transformed into an equivalent minimization problem, namely minimizing the "resilience deficit," as expressed by the following formula: ; In the formula: SOC(t) is the state of charge of the electrochemical energy storage system at time t; α and β are weighting coefficients, and α+β=1; This refers to the rated capacity of the electrochemical energy storage system.

[0069] This design encourages the algorithm to maintain a high average energy storage level and avoid deep discharge during normal times, thereby providing a longer backup time in the event of a green power outage.

[0070] To efficiently solve the aforementioned multi-objective optimization problems involving continuous and discrete variables, high dimensionality, nonlinearity, and strong constraints, this embodiment proposes a swarm adaptive cooperative algorithm (IES-CACA) for integrated energy system optimization. The core of this algorithm lies in constructing multiple co-evolving subpopulations, each focusing on optimizing a specific objective. Through adaptive evolution, inter-population information exchange, and elite archiving mechanisms, the algorithm collaboratively searches for the Pareto optimal frontier.

[0071] In step 3, the Residential Adaptive Cooperative Algorithm (IES-CACA) is used to optimize and solve the constructed multi-objective function, including the following steps: Step 31, Algorithm Initialization and Encoding: Construct a corresponding number of subpopulations and an external archive based on the number of objectives of the multi-objective function. Each subpopulation is used to optimize one objective. The initial subpopulation is generated based on a hybrid initialization strategy of historical scene matching. In this embodiment, based on the target quantity M=3, three subpopulations P1, P2, P3 and an external archive A are initialized. Each subpopulation Pm contains Ns individuals. Each individual corresponds to a complete set of device operation scheduling schemes in the system, which can be encoded using real numbers. The decision variable vector x is encoded as follows: ; Each variable has its upper and lower bounds determined according to the corresponding device constraints; in, This represents the power (kW) purchased through the green electricity direct connection channel during time period t, which directly affects economic targets and is constrained by the upstream green electricity supply capacity. This indicates the electrical power consumed by the deep ground source heat pump during time period t, which must meet the constraints of minimum technical output and power change rate. This indicates the electrical power consumed by the air conditioner during time period t; The operating power of the electrochemical energy storage system during time period t includes charging power and discharging power, and must meet the upper and lower limits of state of charge (SOC) and charging and discharging efficiency constraints. Furthermore, the process of generating the initial subpopulation based on a hybrid initialization strategy using historical scene matching includes the following steps: Step 311: Obtain historical scene data and extract historical scene features; Optionally, historical scenario data can be obtained by building a knowledge base to store historical scenario information and historical optimal scheduling schemes. This knowledge base can be built through offline training and simulation before the system is put into operation. Specifically, historical scene characteristics can include weather coding characteristics, date coding characteristics, historical average electricity load, average proportion of renewable energy output, and other information. The historical scene feature vector can be defined as follows: ; in, Encode the weather type (e.g., 1: sunny, 2: cloudy, 3: rainy). Encode the date type (1: weekday, 2: weekend, 3: public holiday). This represents the historical average electrical load (kW) for this scenario. This represents the average percentage of historical renewable energy output to total load in this scenario.

[0072] Step 312: Extract the features of the current scene to obtain the feature vector of the current scheduling cycle. ; Specifically, the feature vector of the current scene can be the same as that of the historical scene, both including weather coding features, date coding features, historical average electricity load, average proportion of renewable energy output, and other information. Based on the prediction results from step 1, calculate the feature vector for the current scheduling period. ;in, and The average value within the forecast period is used.

[0073] Step 313: Calculate the similarity between the features of the current scene and the features of the historical scenes, obtain the set number of historical scenes that are most similar, and obtain the set of similar scenes; This step involves similar scene matching; specifically, it calculates... With the feature vector of the i-th historical scene Weighted Euclidean distance : ; in, The weight of the j-th feature ( The setting is typically based on the importance of features to scheduling. The K closest historical scenes (e.g., K=3) are selected as the set of similar scenes. .

[0074] Step 314: Hybrid initialization generates a subpopulation, including knowledge-guided individuals generated based on similar scenarios and targeting the corresponding optimization objectives, as well as randomly generated individuals. The knowledge-guided individuals and randomly generated individuals constitute the initialized subpopulation after memory constraint verification. Specifically, for each subpopulation (m=1,2,3), the initial population consists of two parts: Knowledge-guided individuals (Nk): from a set of similar scenarios The historical best solution for each similar scenario In this process, gene fragments optimized for the m-th objective (i.e., partial decision variables) are extracted and Gaussian perturbations are applied to generate new individuals as knowledge-guided individuals. The perturbation formula is: ; in, It is a diagonal covariance matrix. The allowed range of variation for the j-th decision variable is 5% to 10%.

[0075] For example, for the goal of minimizing operating costs F1, a scheduling scheme that affects economic operation is selected as a partial decision variable. The selected partial decision variables are fixed, and perturbations are added to generate multiple individuals. Randomly generated individuals (Nr): Multiple random individuals are randomly generated within the feasible region. This satisfies... ,and The empirical coefficient ρ is 0.3 to 0.5.

[0076] The constraint handling function ensures that all initial individuals satisfy the operational constraints. External archive A is initialized as an empty set with a maximum capacity of N. A .

[0077] Step 32, Adaptive Co-evolution: Based on the optimization contribution of each subpopulation and the predicted uncertainty information, dynamically allocate evolutionary opportunities to guide the mutation operation of individuals, including the following steps: In each generation G, evolutionary opportunities are dynamically allocated and search behavior is adjusted not only based on the recent optimization contribution of the subpopulation, but also coupled with the uncertainty information of source-load prediction.

[0078] Step 321, Quantification of Prediction Uncertainty: Based on the load forecast output from the TFT model in Step 1 and the quantile forecast results of the upstream green electricity supply, the uncertainty index of the forecast for each time period t within the dispatch cycle is obtained. ; Step 322: Calculate the evolutionary probability based on the uncertainty index and the distance traveled. subpopulation The original probability of obtaining an evolutionary opportunity in generation G. Calculated based on recent advance distance, and with an uncertainty adjustment factor introduced. , ; in, This represents the set of key time periods strongly correlated with the optimization objective m, for example, for a resilience objective. , It can include periods when the forecast output of wind and solar power drops sharply; This is the uncertainty sensitivity coefficient (usually taken as 0.5~1.0).

[0079] Evolutionary probability adjusted based on uncertainty adjustment factor for: ; Where m represents the total number of optimization objectives; Step 323, Uncertainty-oriented mutation operation: Screen for high uncertainty periods, calculate the enhanced mutation factor for the screened high uncertainty periods, and perform mutation operation to update individuals; perform individual forward distance calculation, population forward distance calculation and roulette wheel selection on the mutated population to obtain the updated population; For subpopulations Individuals selected for evolution Its mutation vector The generation is based on the original text, and the dimensions of decision variables corresponding to periods of high uncertainty are enhanced.

[0080] set up A set of indices for all decision variable dimensions. For the corresponding The set of decision variable dimensions related to the time period, among which This represents the uncertainty threshold. Enhanced variants Apply by dimension: ; in, Let j be the time period corresponding to dimension j, and η be the coefficient of variation enhancement; subsequently, using Replace the original variant factor Perform mutation operations.

[0081] Subsequent steps such as calculating individual progress distance, population progress distance, and roulette wheel selection remain unchanged, but the adjusted evolutionary probabilities are used. The algorithm is selected and mutated using an enhanced mutation factor. This step quantifies uncertainty, adjusts evolutionary resources, and enhances mutation in key dimensions, enabling the algorithm to efficiently find the best solution even under fluctuating source load scenarios, thus balancing convergence speed and solution robustness.

[0082] Traditional algorithms are prone to getting trapped in local optima due to the high-dimensional and nonlinear characteristics of integrated energy system optimization models for office buildings, while metaheuristic algorithms struggle to balance convergence speed and solution accuracy. IES-CACA improves global optimization capabilities under complex constraints through multi-subpopulation co-evolution (each subpopulation focuses on one objective) and adaptive evolutionary strategies (dynamically allocating evolutionary opportunities).

[0083] Step 33, Subpopulation migration strategy: Every set number of iterations, such as every R generations, first evaluate the subpopulation status (calculate the standardized median and interquartile range) and pair superior and inferior subpopulations based on comprehensive ranking. Then, exchange information through a two-way individual migration method that transfers superior genes from superior populations and introduces diversity from inferior populations to obtain the population after the individual migration operation. The population size remains unchanged after migration. Every R generations (e.g., R=10), a migration operation is performed to facilitate information exchange between targets and prevent premature convergence of the search.

[0084] Step 33, the method for generating the population after individual migration operations, includes the following steps: 331) Subpopulation Status Assessment: For each subpopulation Pm, sort all individuals in subpopulation Pm in ascending order according to the m-th target value fm. Calculate its standardized median. and standardized interquartile range : ; in, Let be the minimum, first quartile, median, third quartile, and maximum value of the m-th target value, respectively.

[0085] 332) Calculate the overall ranking of each subpopulation, perform community cooperative pairing, and pair the best and worst, and the second best and second worst according to the ranking value. ; in, Indicates ascending order (med) m Smaller is better) This indicates a descending ranking (a higher IQRm indicates better diversity).

[0086] Pair the subpopulation with the lowest rank value with the subpopulation with the highest rank value, then pair the next best with the next worst, and so on, to achieve collaborative optimization among communities.

[0087] 333) Individual migration: For a pair of superior populations Pi and inferior populations Pj, the superior individuals in the superior population Pi are migrated to the inferior population Pj; after the migration, the same number of individuals in the inferior population Pj are migrated to the superior population Pi. Specifically, the superior subpopulation Pi migrates to the inferior subpopulation Pj: individuals in subpopulation Pi whose target value i is greater than its third quartile Q3j are migrated to Pj; Migration from inferior subpopulation Pj to superior subpopulation Pi: an equal number of individuals in subpopulation Pj whose value on the i-th target is greater than its first quartile Q1i are migrated to Pi. After the migration is completed, both subpopulations accept the new individuals and randomly remove some of the original individuals to keep the population size Ns unchanged.

[0088] Traditional single-objective optimization neglects the synergy between economy, stability, and resilience, while multi-objective optimization involves conflicts between objectives (such as cost and resilience). IES-CACA promotes information exchange between objectives through a subpopulation migration strategy (pairing and exchanging individuals based on convergence and diversity), generating a Pareto optimal solution set that covers the trade-offs between multiple objectives, providing rich choices for dynamic decision-making.

[0089] Step 34, External Archive Update and Maintenance Strategy: After each generation of evolution, all individuals are merged to form a candidate solution set. Non-dominated solutions are selected by non-dominated sorting. If the number exceeds the archive capacity NA, redundant solutions are pruned by target value normalization, angle selection of similar solution pairs, and offset density estimation (SDE). After each generation of evolution, the gregarious adaptive cooperative algorithm merges all individuals from the current three subpopulations, individuals from the external archive A, and new individuals generated through differential evolution to form a candidate solution set. Non-dominated solutions are then selected through non-dominated sorting. If the number of non-dominated solutions exceeds the archive capacity NA, a pruning strategy based on angle selection and offset density estimation is implemented.

[0090] Objective value normalization: Let the ideal point of the current candidate solution set in the objective space be: ; The worst result is: ; For the solution Its normalized objective value vector The m-th component is: ; Angle selection: Calculate any two solutions and The angle between the normalized target vectors : ; Find the included angle of all solutions The smallest pair of solutions ( , These represent the search directions that are most similar.

[0091] Offset density estimation: for the found similar solution pairs ( , ), calculate its offset density estimate to determine the removal item.

[0092] To solve For example, for any other solution Calculate its relative to offset target vector Its m-th component is defined as:

[0093] This operation will target those that are not superior in all objectives. The solution is "offset" to The position, thus orbiting in the target space. This forms a solution set after the "offset".

[0094] Then calculate Its k (usually) The "offset" distance of the nearest neighbors. Solution Offset density estimate Defined as: ; in, yes The "offset" distance to its k-th nearest neighbor.

[0095] Cutting Decisions: Comparison and The SDE value is used to remove solutions with larger SDE values ​​(i.e., those located in more crowded areas or with worse overall performance). The process of "angle selection - offset density estimation - pruning" is repeated until the archive size is restored to NA.

[0096] Traditional archive update strategies struggle to maintain the distribution and convergence of Pareto optimal solutions, resulting in a limited coverage of optimization schemes. This algorithm, through an archive update strategy based on angle selection and offset density estimation, prioritizes retaining elite solutions with significant differences in search directions while eliminating redundant solutions, achieving a balance between solution set diversity and convergence.

[0097] Step 35, Iteration Termination and Output of Optimal Scheduling Scheme Set: Iterate through steps 32 to 34 until the iteration termination condition is met, obtaining the Pareto optimal scheduling scheme set, which is output through an external archive A; this archive contains NA Pareto optimal scheduling schemes, each scheme Xj containing three objective function values ​​(total operating cost F). 1,j Equipment power fluctuation F 2,j System power supply resilience F 3,j and a complete equipment power scheduling sequence; The iteration cutoff condition can be: reaching the preset maximum number of generations Gmax, or the improvement of the Hypervolume metric of the external archive A within K consecutive generations being less than a threshold. The algorithm ultimately outputs an external archive A, which contains a set of Pareto optimal scheduling schemes that do not dominate each other on the three objectives {F1, F2, F3}. Each solution represents a specific trade-off between operating costs, equipment stability, and system resilience, providing a wealth of optimization options for subsequent decision-making.

[0098] Further technical solutions include step 5 after step 3: for the obtained Pareto optimal scheduling scheme set, based on dynamic preference-driven and multi-attribute collaborative projection decision, select the optimal scheduling scheme to regulate the operation of the integrated energy system. This step aims to automatically select the optimal compromise scheduling scheme that best matches the current system operating state from the Pareto optimal scheduling scheme set output by the IES-CACA algorithm. This invention proposes a dynamic preference-driven and multi-attribute collaborative projection decision-making method, which adaptively generates decision preferences by sensing the real-time system state and comprehensively evaluates the overall performance of each scheme to finally determine the execution scheme. The specific process is as follows, based on the dynamic preference-driven and multi-attribute collaborative projection decision-making process, including the following steps: S51. Obtain the Pareto optimal scheduling scheme set and construct the candidate scheduling scheme set; Receive an external archive A from the IES-CACA algorithm optimization output in S4. This archive contains NA Pareto optimal scheduling schemes, forming a candidate scheme set. Each scheme Xj corresponds to three objective function values: total operating cost F1,j, equipment power fluctuation F2,j, and system power supply resilience F3,j, and includes a complete equipment power scheduling sequence.

[0099] S52. Dynamic Preference Generation: Generate weight vectors for different objectives based on the acquired state features; Step 521: Obtain the current system state information and extract the current system state feature vector: ; Wherein, s1 is the normalized value for the current period (0-23 hours); s2 is the current state of charge (SOC) (t0) of the electrochemical energy storage; and s3 is the predicted fluctuation coefficient of wind and solar power output for the next 24 hours. ;s4 represents the extreme weather warning level (0-1, 0 indicates no warning);s5 represents the real-time carbon price coefficient (normalized).

[0100] Step 522: Construct a dynamic preference generation network. Input the state vector into the pre-trained lightweight dynamic preference generation network, and output a dynamic preference weight vector for the three optimization objectives. The specific calculation process is as follows: ; In the formula, W1, b1, W2, and b2 are network parameters; ReLU(.) is the linear rectified activation function; Softmax(.) is the normalized exponential function, ensuring that w1 + w2 + w3 = 1 and wi > 0. The weights w1, w2, and w3 represent the degree of importance attached to economy (F1), stability (F2), and resilience (F3) in the current state, respectively.

[0101] The dynamic preference generation network employs a deep learning network. S53. Multi-attribute collaborative projection and comprehensive scoring: Combining dynamic preference weights, the basic fitness, balance penalty (Imbalancej), and extreme solution penalty (Extremej) of the scheduling scheme are calculated, and the comprehensive score is finally obtained. The three objective values ​​of each scheme Xj in the candidate scheme set are normalized, and its dynamic comprehensive score is calculated.

[0102] First, calculate the minimum value of each objective. and maximum value Perform linear normalization: ; Obtain the normalized target vector ; Next, the basic fitness, balance penalty, and overall score of scheme Xj under dynamic preferences are calculated as follows: ; In the formula, Imbalance j To account for the bias-weighted objective imbalance; λ is the equilibrium penalty coefficient, which is adaptively adjusted according to the system state: ; Among them, λ0=0.5, α=0.3, β=0.4; Extreme j This is a penalty term for extreme solutions, used to penalize schemes with excessively large differences in target values.

[0103] Where τ=0.7; γ=2.0 are the penalty intensity.

[0104] S54. Interactive Fine-tuning and Determination of Optimal Solution: The final optimal scheduling solution is determined based on the comprehensive score and manual fine-tuning. The system automatically recommends the solution with the highest overall score. This serves as the initial recommended optimal compromise scheduling scheme. Simultaneously, the human-machine interface of the energy management system displays the current dynamic preference weight *w* and the comprehensive score ranking of candidate schemes, allowing operators to fine-tune the preference weight using a slider (e.g., manually increasing *w*3 during extreme weather warnings). The system responds in real-time to adjustments, recalculating and refreshing the recommended scheme. Finally, after confirmation or fine-tuning by the operators, the selected scheme is finalized. The output is the final optimal compromise scheduling scheme, which contains hourly power instructions for each device and will be sent to each local controller for execution.

[0105] Furthermore, the optimal scheduling scheme is output, specifically the optimal compromise scheduling scheme finally determined by the optimization method, and transformed into a set of control commands that can be issued. This scheme is the final embodiment of the aforementioned prediction, modeling, optimization, and decision-making steps, and will be used to directly guide the real-time operation of the integrated energy system of the office building during the new scheduling cycle.

[0106] Specifically, the output optimal compromise scheduling scheme Includes the following complete information: 1) The hourly real-time power distribution sequence of each controllable device, i.e., the electrical power of the medium-deep ground source heat pump. Electric power of air conditioner The charging and discharging power of electrochemical energy storage systems and the power purchased from the green electricity direct connection channel These sequences together constitute the direct power reference for system operation; 2) The operating state curve of the energy storage system, i.e., the planned state of charge of electrochemical energy storage during the dispatch cycle. The curve visually reflects the energy planning that the system has made to mitigate fluctuations and ensure resilience. 3) The numerical values ​​achieved by the three optimization objectives, i.e., the total operating cost corresponding to this solution. Equipment power fluctuation index Quantitative value of system power supply resilience This is used to evaluate the overall performance of the solution; The interaction information between the integrated energy supply system directly connected to green electricity and the external network mainly includes the planned power purchase sequence with upstream wind and solar industrial parks, and the power exchange with the power grid (backup) (if any). This information will be decomposed and distributed to the corresponding field execution units such as ground source heat pump controllers, electric air conditioning group control systems, and energy storage converters (PCS) through the communication network of the energy management system. At the same time, it will be visualized on the human-machine interface, thereby completing the closed loop from intelligent decision-making to physical execution, and realizing the safe, economical, stable and highly resilient operation of the integrated energy system of office buildings under the green electricity direct connection mode.

[0107] Example 2 Based on Example 1, this example provides a green electricity direct-connection energy supply optimization system that considers uncertainties and system resilience, including: The prediction module is configured to build a time-fusion Transformer model, obtain the operating data of the green electricity direct-connection power supply system to perform load prediction and upstream green electricity supply prediction, and calculate uncertainty; The model building module is configured to construct a multi-objective optimization model with the optimization objectives of minimizing total operating cost, minimizing equipment power fluctuations, and maximizing system energy supply resilience. The solution module is configured to use a swarm adaptive cooperative algorithm to solve for the Pareto optimal scheduling scheme set for a multi-objective optimization model, so as to regulate the operation and scheduling of the green electricity direct-connection power supply system. In the process of solving the problem using the gregarious adaptive cooperative algorithm, during the initialization phase, a corresponding number of subpopulations and an external archive are constructed according to the number of objectives in the multi-objective optimization model. Each subpopulation is used to optimize one objective. An initial subpopulation is generated based on a hybrid initialization strategy that matches historical scenarios. During the population evolution phase, an adaptive evolution strategy is adopted, and evolutionary opportunities are dynamically allocated based on the optimization contribution of each subpopulation and the uncertainty information of the prediction.

[0108] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0109] Example 3 Based on Example 1, this example provides an optimized green electricity direct-connection power supply system that considers uncertainties and system resilience, such as... Figure 3 As shown, a green direct-connection energy supply system for office buildings is constructed, including: an upstream wind and solar industrial park, photovoltaic panels, lithium iron phosphate batteries, and a power management system. The integrated energy system includes ground source heat pumps and electric air conditioners, etc., to provide end loads, including electrical loads, heat loads, and cooling loads. Among them, the power management system, as the core control unit of the system, is configured to execute the green direct-connection energy supply optimization method considering uncertainty and system resilience described in Example 1.

[0110] Example 4 Based on Embodiment 1, this embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps in the green direct-connection energy supply optimization method considering uncertainty and system resilience described in Embodiment 1.

[0111] Example 5 Based on Embodiment 1, this embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps in the green direct-connection energy supply optimization method considering uncertainty and system resilience described in Embodiment 1.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0113] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this 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 without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An optimization method for green electricity direct connection energy supply considering uncertainty and system resilience, characterized in that, Includes the following steps: A time-fusion Transformer model is constructed to obtain the operating data of the green electricity direct-connection power supply system for load forecasting and upstream green electricity supply forecasting, and to calculate the uncertainty. A multi-objective optimization model is constructed with the optimization objectives of minimizing total operating cost, minimizing equipment power fluctuations, and maximizing system energy supply resilience. For the multi-objective optimization model, the swarm adaptive cooperative algorithm is used to solve the Pareto optimal scheduling scheme set to regulate the operation and scheduling of the green electricity direct-connection power supply system. In the process of solving the problem using the gregarious adaptive cooperative algorithm, during the initialization phase, a corresponding number of subpopulations and an external archive are constructed according to the number of objectives in the multi-objective optimization model. Each subpopulation is used to optimize one objective. An initial subpopulation is generated based on a hybrid initialization strategy that matches historical scenarios. During the population evolution phase, an adaptive evolution strategy is adopted, and evolutionary opportunities are dynamically allocated based on the optimization contribution of each subpopulation and the uncertainty information of the prediction.

2. The green electricity direct-connection energy supply optimization method considering uncertainty and system resilience as described in claim 1, characterized in that: A method for constructing a time-fusion Transformer model to predict load and upstream green electricity supply, and to calculate uncertainties, includes the following steps: The data is cleaned and normalized, and then the processed data is divided into covariates to obtain static covariates, known future covariates, and historical observation covariates. For static covariates, a static encoder consisting of multiple GRNs is used to encode the static covariates into multiple context vectors; For historical observation covariates and known future covariates, GRN transformation, feature concatenation, weight calculation and weighting are performed sequentially to obtain the final historical feature representation and known future feature representation. The processed historical feature representation is input into the sequence encoder for temporal feature extraction to obtain the encoded features; For the encoded features, perform an interpretable multi-head attention operation, perform one-dimensional linear transformation and temporal fusion, and fuse with known future features to achieve decoding to obtain the hidden state of each future time step; The hidden state of each future time step is obtained and mapped to the final predicted value through a fully connected output layer, thereby obtaining the quantile results of the predicted load and the predicted green power output for each time period. Based on the quantile results of different prediction targets for each time period, the uncertainty index is calculated.

3. The green electricity direct-connection energy supply optimization method considering uncertainty and system resilience as described in claim 1, characterized in that: For the constructed multi-objective function, a swarm adaptive cooperative algorithm is used for optimization and solution, including the following steps: Based on the number of objectives of the multi-objective function, a corresponding number of subpopulations and an external archive are constructed. Each subpopulation is used to optimize one objective. An initial subpopulation is generated based on a hybrid initialization strategy of historical scene matching. Based on the optimization contribution of each subpopulation and the predicted uncertainty information, evolutionary opportunities are dynamically allocated to guide the mutation operations of individuals. Every set number of iterations, the subpopulation status is first evaluated and the superior and inferior subpopulations are paired based on a comprehensive ranking. Then, information is exchanged through a two-way individual migration method that transmits superior genes to the superior population and introduces diversity to the inferior population, resulting in the population after the individual migration operation. After each generation of evolution, all individuals are merged to form a candidate solution set, and non-dominated solutions are selected by non-dominated sorting. Iterate through the above steps until the iteration deadline is met to obtain the Pareto optimal scheduling scheme set, which is then output through an external archive.

4. The green electricity direct-connection energy supply optimization method considering uncertainty and system resilience as described in claim 3, characterized in that: The process of generating an initial subpopulation based on a hybrid initialization strategy using historical scene matching includes the following steps: Acquire historical scene data and extract historical scene features; Extract the features of the current scenario to obtain the feature vector of the current scheduling cycle; Calculate the similarity between the features of the current scene and the features of historical scenes, obtain the set of most similar historical scenes, and obtain the set of similar scenes; The hybrid initialization generates a subpopulation, which includes knowledge-guided individuals generated based on similar scenarios and targeting corresponding optimization objectives, as well as randomly generated individuals. The knowledge-guided individuals and randomly generated individuals are constrained and verified to form the initialization subpopulation.

5. The green electricity direct-connection energy supply optimization method considering uncertainty and system resilience as described in claim 3, characterized in that: Based on the optimization contribution of each subpopulation and the predicted uncertainty information, evolutionary opportunities are dynamically allocated to guide the mutation operations of individuals, including the following steps: Based on the quantile forecast results of load forecasting and upstream green electricity supply forecasting, the forecast uncertainty index for each period within the dispatch cycle is obtained. Calculate the evolutionary probability based on the uncertainty index and the distance traveled: High uncertainty periods are screened, and enhanced mutation factors are calculated for the screened high uncertainty periods. Mutation operations are then performed on the individuals to update them. Perform individual forward distance calculation, population forward distance calculation, and roulette wheel selection on the mutated population to obtain the updated population.

6. The green electricity direct-connection energy supply optimization method considering uncertainty and system resilience as described in claim 3, characterized in that: The method for generating a population after individual migration operations includes the following steps: For each subpopulation, arrange all individuals in the subpopulation in ascending order of the target and calculate the standardized median and standardized interquartile range. Calculate the overall ranking of each subpopulation, perform community cooperative pairing, and pair the best and worst, and the second best and second worst, according to the ranking value. For a pair of superior offspring populations Pi and inferior offspring populations Pj, the superior individuals in the superior offspring population Pi are transferred to the inferior offspring population Pj, and after the transfer, the same number of individuals in the inferior offspring population Pj are transferred to the superior offspring population Pi.

7. The green electricity direct-connection energy supply optimization method considering uncertainty and system resilience as described in claim 1, characterized in that: For the obtained Pareto optimal scheduling scheme set, based on dynamic preference-driven and multi-attribute cooperative projection decision-making, the optimal scheduling scheme is selected to regulate the operation of the integrated energy system; the dynamic preference-driven and multi-attribute cooperative projection decision-making process includes the following steps: Obtain the Pareto optimal scheduling scheme set and construct a candidate scheduling scheme set; Generate weight vectors for different objectives based on the acquired state features; By combining dynamic preference weights, the basic fitness, balance penalty, and extreme solution penalty of the scheduling scheme are calculated, and the comprehensive score is finally obtained. The final optimal scheduling scheme is determined based on the overall score and manual fine-tuning.

8. A green electricity direct-connection energy supply optimization system considering uncertainty and system resilience, characterized in that, include: The prediction module is configured to build a time-fusion Transformer model, obtain the operating data of the green electricity direct-connection power supply system to perform load prediction and upstream green electricity supply prediction, and calculate uncertainty; The model building module is configured to construct a multi-objective optimization model with the optimization objectives of minimizing total operating cost, minimizing equipment power fluctuations, and maximizing system energy supply resilience. The solution module is configured to use a swarm adaptive cooperative algorithm to solve for the Pareto optimal scheduling scheme set for a multi-objective optimization model, so as to regulate the operation and scheduling of the green electricity direct-connection power supply system. In the process of solving the problem using the gregarious adaptive cooperative algorithm, during the initialization phase, a corresponding number of subpopulations and an external archive are constructed according to the number of objectives in the multi-objective optimization model. Each subpopulation is used to optimize one objective. An initial subpopulation is generated based on a hybrid initialization strategy that matches historical scenarios. During the population evolution phase, an adaptive evolution strategy is adopted, and evolutionary opportunities are dynamically allocated based on the optimization contribution of each subpopulation and the uncertainty information of the prediction.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the green direct-connection energy supply optimization method considering uncertainty and system resilience as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the steps in the green direct-connection energy supply optimization method considering uncertainty and system resilience as described in any one of claims 1-7.

Images