Zero-carbon building energy system collaborative optimization control method

Abstract

The invention relates to the technical field of zero-carbon building energy system optimization, and discloses a zero-carbon building energy system collaborative optimization control method, which comprises a multi-time step data hierarchy construction method. According to the zero-carbon building energy system collaborative optimization control method, a multi-time-step data hierarchy construction method is adopted, and the feature dimension of data is greatly enriched. Specifically, according to the method, original 15-minute-granularity data is dynamically aggregated into 30-minute and 1-hour-level data, aggregation is carried out in a weighted average mode, so that information of different time scales is reserved and integrated, and the data of each time level can reflect the change rule of energy data under different time scales. For example, instant fluctuation of energy consumption can be captured through 15-minute hierarchical data, and the one-hour hierarchical data is more suitable for displaying the overall trend of energy consumption, so that the model can learn short-term fluctuation and long-term trend at the same time through the multi-time-step construction mode.

Description

Technical Field

[0001] The present invention relates to the technical field of zero-carbon building energy system optimization, and particularly to a collaborative optimization control method for a zero-carbon building energy system. Background Technique

[0002] A zero-carbon building refers to a building that, during its entire life cycle, through the adoption of energy-saving technologies, renewable energy utilization, carbon offset and other measures, makes the net carbon emissions of the building approach zero. Its core goal is to achieve harmonious coexistence between the building and the natural environment, reduce dependence on fossil energy, lower greenhouse gas emissions, and contribute to addressing climate change. With the intensification of global climate change and the full promotion of the carbon neutrality goal, as the core carrier of the low-carbon transformation in the building field, the optimization control of the energy system of zero-carbon buildings has become the focus of common concern in the academic and industrial circles.

[0003] Traditional zero-carbon building energy systems generally have bottleneck problems such as high dependence on fossil energy, low multi-energy coupling efficiency, and lag in dynamic response of supply and demand. At the same time, the energy management system mostly uses a single time resolution (such as 1 hour) for energy consumption prediction and scheduling. Since it cannot capture short-term (15-minute level) fluctuation characteristics such as photovoltaic power output and heat pump start-stop, it leads to a lag in scheduling response. At the same time, there is a lack of collaborative analysis of the energy supply and demand coupling relationship at different time scales (such as the minute-level charge and discharge of energy storage and the hourly load demand). Moreover, single-time-step data is vulnerable to abnormal interferences such as sudden weather changes and equipment failures, resulting in significant accumulation of prediction errors. At the same time, the high-frequency noise of short-time-step data interferes with the scheduling stability, the lag of long-time-step data reduces the new energy consumption capacity, and the lack of cross-time-step association rules leads to the failure of subsystem collaboration. This way, the real-time performance and robustness of the zero-carbon building energy system are reduced. Therefore, a collaborative optimization control method for a zero-carbon building energy system is proposed. Summary of the Invention

[0004] (I) Technical Problems to be Solved In view of the deficiencies of the prior art, the present invention provides a collaborative optimization control method for a zero-carbon building energy system to solve the technical problems of the reduced real-time performance and robustness of the zero-carbon building energy system described above.

[0005] (II) Technical Solutions To achieve the above object, the present invention provides the following technical solutions: A collaborative optimization control method for a zero-carbon building energy system includes the following steps: Step 1, collect multi-source data of the building energy system in real time, including meteorological data, energy consumption data, and equipment operation status data; Step 2: Perform multi-time-step decomposition on the original 15-minute granularity data to construct a data pyramid containing three levels of 15 minutes, 30 minutes, and 1 hour. The data between levels is weighted and aggregated through dynamic weight coefficients α and γ; Step 3: Use the sliding window technique to extract local features from the data at each level, construct a three-branch LSTM network to process multi-time-step features respectively, and achieve multi-scale feature fusion through the attention mechanism; Step 4: Establish an optimization model with the dual objectives of minimizing the operating cost and maximizing the renewable energy consumption rate, and set power balance constraints, equipment operation constraints, energy storage SOC constraints, and carbon emission constraints; Step 5: Use a mixed-integer linear programming solver for rolling optimization control, and update the scheduling plan for the next 4 hours every 15 minutes; Step 6: Real-time verify the prediction error and trigger model retraining, and dynamically update the feature weights and optimization parameters.

[0006] Preferably, the data aggregation between levels in Step 2 adopts the following formula: The 15-minute time step is level 1: The time series is minutes, corresponding to 96 data points, denoted as ; The 30-minute time step is level 2: Generated by aggregating the data at level 1, the time series is minutes, corresponding to 48 data points, denoted as ; The 1-hour time step is level 3: Generated by aggregating the data at level 2, the time series is minutes, corresponding to 24 data points, denoted as ; Each data point at level 2 is obtained by the weighted average of two adjacent data points at level 1: where is the time weight coefficient: When it is an equal-weight average; When it emphasizes the data in the first 15 minutes; The data point at level 3 is aggregated from the data points at level 2:

[0007] where is the weight coefficient.

[0008] Preferably, the local feature extraction formula in step three is as follows:

[0009] where n is the window length and m is the prediction step size.

[0010] Preferably, the attention mechanism is calculated using the following formula:

[0011] where is the attention weight of the k-th level feature, is the learnable weight matrix, is the output feature of the k-th level LSTM network; The fused feature is .

[0012] Preferably, the dual-objective optimization function in step four specifically includes: Economic objective:

[0013] where each cost item is calculated according to the time-of-use electricity price, gas price, and carbon price; Environmental protection objective:

[0014] where the calculation of the renewable energy consumption rate includes a discarded power correction term.

[0015] Preferably, the constraint conditions set in step four specifically include: Power balance equation: ; Equipment operation constraint: ; ; Energy storage SOC dynamic model: ; .

[0016] Preferably, the specific implementation method of the rolling optimization strategy in step five is as follows: Each optimization covers the next 16 15-minute time steps, and the optimization variables include the grid interaction power, gas equipment power, energy storage charge and discharge status, and surplus power grid connection power; The branch and bound algorithm is used to solve the MILP problem, and the solution time window is restricted within 3 minutes.

[0017] Preferably, the feedback mechanism setting in step six includes: Calculate the RMSE value of the predicted results for the next 1 hour in real time. When the RMSE > 5% of the rated load for three consecutive time steps, trigger model retraining; During the retraining process, the online gradient descent algorithm is used to update the attention weight matrix W_k and the LSTM network parameters.

[0018] Preferably, the dynamic weight adjustment strategy specifically includes adjusting according to the real-time light intensity and , using the following formula: ; .

[0019] The specific principle of this solution is as follows: 1. The specific principle of multi-time-step data decomposition: Multi-time-step data decomposition can effectively capture the dynamic characteristics of the energy system at different time scales by gradually aggregating the original high-resolution data (such as 15-minute time steps) into multi-granularity levels such as 30 minutes and 1 hour. This method breaks through the limitations of traditional single-step analysis through data hierarchical processing, providing a key data basis for the multi-energy collaborative optimization of zero-carbon buildings.

[0020] Divide 24 hours of a day into 3 time granularity levels to construct a multi-resolution data pyramid: Level 1 (15-minute time step): The time series is minutes, corresponding to 96 data points, denoted as .

[0021] Level 2 (30-minute time step): Generated by aggregating the Level 1 data, the time series is minutes, corresponding to 48 data points, denoted as .

[0022] Level 3 (1-hour time step): Generated by aggregating the Level 2 data, the time series is minutes, corresponding to 24 data points, denoted as .

[0023] Each data point in Level 2 is obtained by the weighted average of two adjacent data points in Level 1: where is the time weight coefficient. When it is an equal-weight average; when it emphasizes the data of the first 15 minutes. For example, for time-sensitive energy (such as electric load), if the data fluctuates violently during a certain period (such as the morning peak), increase ( > 0.5) can assign higher weights to the first 15 minutes, reflecting the timeliness of neighboring data. If the data changes smoothly (such as at night), take = 0.5 to achieve smoothing and noise reduction. Here, it can be optimized in real time through an online learning algorithm (such as exponentially weighted moving average) value.

[0024] Similarly, the level 3 data points are aggregated from the level 2 data points: where is the weight coefficient, which can be dynamically adjusted according to the energy type (for example, photovoltaic power generation emphasizes the daytime period). For time-dependent energy sources (photovoltaic power generation): The light changes rapidly during the day, and β is dynamically allocated in different time periods: sunrise / sunset period: β > 0.5, emphasizing the data in the first half hour (capturing rapid changes). Noon stable period: β = 0.5 (smooth averaging).

[0025] 2. Specific principle of feature extraction: By aggregating the original 15-minute granularity data into 30-minute and 1-hour granularity data step by step and combining the sliding window technique, it is possible to capture high-frequency load fluctuations (such as minute-level air conditioner startups and shutdowns), medium-frequency trend changes (such as morning and evening electricity consumption peaks), and low-frequency periodic patterns (such as weekday / weekend energy consumption patterns) simultaneously.

[0026] For each level of data , local features are extracted using a sliding window:

[0027] where n is the window length and m is the prediction step. For example, take n = 24 (1 day) and m = 4 (1 hour).

[0028] The multi-scale feature fusion model dynamically allocates the feature weights of each level through the attention mechanism, effectively solving the problem of the traditional single-step model's lag in response to sudden changes in energy supply and demand, significantly improving the prediction accuracy, and providing decision-making basis for the coordinated scheduling of energy storage and distributed power sources in multiple time dimensions, enhancing the system's adaptability to the randomness of renewable energy output.

[0029] 3. Specific principle of attention mechanism fusion: The attention mechanism plays a key role in the fusion of multi-time step data, and realizes the adaptive integration of multi-scale information by dynamically calculating the importance weights of each level of features.

[0030] Introduce the self-attention mechanism to calculate the importance weights of each level of features :

[0031] where is a learnable weight matrix.

[0032] This mechanism is based on the principle of self-attention and automatically determines the contribution degree of features at different time granularities (15 minutes, 30 minutes, 1 hour) to the prediction target through a learnable weight matrix. This dynamic allocation strategy breaks through the limitations of traditional fixed-weight fusion, enabling the model to capture the coupling relationship between short-term fluctuations and long-term trends in energy supply and demand, significantly improving prediction accuracy and the adaptability of optimal control.

[0033] 4. Prediction Output The prediction output is obtained by means of multi-time step data decomposition, a feature fusion model, and a collaborative optimization strategy, which can accurately estimate the future energy load of zero-carbon buildings, provide a basis for multi-energy collaborative scheduling, assist in the rational planning of energy supply, and enhance system stability and energy utilization efficiency.

[0034] Map the fused features to the prediction target (such as the future 15-minute load value) through a fully connected layer:

[0035] The loss function uses the mean squared error (MSE): 5. Principle of the Optimization Objective Function Take the minimization of system operation cost and the maximization of renewable energy consumption rate as dual objectives: Economic objective:

[0036] Environmental protection objective:

[0037] Cost item:

[0038]

[0039]

[0040]

[0041] Among them is the power purchased from the power grid, is the time-of-use electricity price; is the power of gas equipment; is the gas carbon emission coefficient; is the carbon price; is the power of surplus electricity fed into the grid.

[0042] Consumption rate item:

[0043] Among them is the output of renewable energy (such as the power generation of photovoltaic and wind power), is the curtailed power (the power of renewable energy not utilized).

[0044] This bi-objective function realizes the coordinated optimization of the economy and environmental protection of the zero-carbon building energy system by minimizing the system operation cost (grid power purchase, gas consumption, carbon emission cost and surplus power income) and maximizing the renewable energy consumption rate.

[0045] 6. Specific principles of constraint conditions The constraint conditions ensure that the multi-energy system operates within a safe and economic range, including equipment power limits, energy storage SOC constraints, user comfort thresholds, etc., avoiding equipment overload or energy storage failure, while ensuring the basic energy demand of users and maintaining the stable and reliable operation of the system.

[0046] Power balance constraint:

[0047] Among them : Grid power purchase (positive value for power purchase, negative value for power sale), : Gas power generation, : Energy storage charge and discharge power (discharge is positive, charge is negative), : Surplus power grid connection power (power sale), : Load demand power, : Curtailed power.

[0048] Equipment operation constraint:

[0049]

[0050] Energy storage SOC constraint:

[0051]

[0052] Carbon emission constraint:

[0053] Among them is the energy storage efficiency, is the time step (15 minutes), is the total capacity of the energy storage system, 、 is the charge / discharge power (both take positive values), Represents the power generation of renewable energy sources (such as solar energy, wind energy, etc.) in a zero-carbon building system, usually measured in kilowatt-hours (kWh). Represents the total energy demand of the building, that is, all the energy consumed during the operation of the building (including renewable energy, electricity purchased from the grid, gas, etc.), measured in kilowatt-hours (kWh).

[0054] (III) Beneficial effects Compared with the prior art, the present invention provides a collaborative optimization control method for a zero-carbon building energy system, having the following beneficial effects: 1. The collaborative optimization control method for the zero-carbon building energy system adopts a multi-time-step data hierarchy construction method, which greatly enriches the feature dimensions of the data. Specifically, this method dynamically aggregates the original 15-minute granular data into 30-minute and 1-hour hierarchical data, and performs the aggregation through weighted averaging, so that the information at different time scales can be retained and integrated. The data at each time level can reflect the variation laws of energy data at different time scales. For example, the 15-minute level data can capture the instantaneous fluctuations of energy consumption, while the 1-hour level data is more suitable for showing the overall trend of energy consumption. This multi-time-step construction method enables the model to simultaneously learn short-term fluctuations and long-term trends, greatly improving the model's understanding ability of the complex changes in energy data. In addition, dynamically adjusting the aggregation weights according to real-time meteorological data further enhances the adaptability of the model to weather-sensitive energy sources (such as photovoltaic), enabling the model to better respond to the changes in energy data under different weather conditions; 2. The collaborative optimization control method for the zero-carbon building energy system effectively improves the model's ability to capture the features of energy data through the multi-scale feature enhancement and fusion strategy. The sliding window is applied to the data at each time level to extract local features, and the data within the window can reflect the variation patterns of energy data within a certain time range. By constructing a three-branch LSTM network to process the data at different levels respectively, the feature information of each level of data is fully mined. At the same time, the attention mechanism is introduced to fuse the multi-scale features, and the learnable weights can automatically allocate the importance of each level of features. This enables the model to more accurately focus on the key features when facing complex energy data, avoiding the information imbalance problem that may be caused by the traditional fixed-weight fusion method. This fusion method not only improves the utilization efficiency of features, but also enhances the adaptability and generalization ability of the model to different types of energy data; 3. The collaborative optimization control method for the zero-carbon building energy system. The construction of the multi-objective optimization model realizes the maximization of the comprehensive benefits of the multi-energy system of the zero-carbon building. Optimizing with the lowest operating cost and the highest renewable energy consumption rate as the dual objectives fully considers the economy and environmental protection of the energy system. During the optimization process, constraint conditions such as power balance, equipment operation limits, and energy storage SOC are set to ensure the safe and stable operation of the energy system. By solving the optimal scheduling plan through mixed-integer linear programming (MILP), an energy scheduling strategy that optimizes the two objective functions can be found under the premise of meeting various constraint conditions. This multi-objective optimization method enables the zero-carbon building to maximize the consumption of renewable energy while reducing the operating cost, meeting the requirements of sustainable development; 4. The collaborative optimization control method for the zero-carbon building energy system. The rolling optimization strategy ensures the real-time regulation and dynamic adaptability of the energy system. Input the latest data every 15 minutes, update the energy supply and demand forecast for the next 4 hours, and re-solve the optimization problem to generate a new scheduling plan. This real-time update and optimization method can respond promptly to various changes in the energy system, such as real-time meteorological changes and equipment failures. By continuously adjusting the energy storage charge and discharge, heat pump operation mode, etc., the energy system is always maintained in the optimal operating state. At the same time, the rolling optimization strategy combines short-term forecasting and real-time regulation, improving the energy system's ability to cope with uncertainties and reducing the operating risks caused by prediction errors and external disturbances; 5. The collaborative optimization control method for the zero-carbon building energy system. The real-time data verification and feedback mechanism ensure the accuracy and reliability of the model. Calculate the root mean square error (RMSE) between the predicted value and the measured value in real time and set an error threshold. When the error exceeds the threshold, the model re-training is automatically triggered to update the feature weights and optimization parameters. This feedback mechanism enables the model to continuously adjust and optimize according to the actual operating conditions, improving the model's ability to track changes in energy data. Through continuous verification and feedback, the model can gradually adapt to the dynamic changes of the energy system, reduce prediction errors, and ensure the efficient operation of the energy system; 6. The collaborative optimization control method for the zero-carbon building energy system. Through a series of innovative methods such as multi-time step feature fusion, multi-objective optimization, and rolling regulation, the present invention comprehensively improves the prediction accuracy and regulation efficiency of the multi-energy system of the zero-carbon building. When dealing with complex and changing energy data, it can more accurately capture key features, realize the efficient allocation and utilization of energy, and significantly improve the economy, environmental protection, and stability of the energy system, providing strong technical support for the development of zero-carbon buildings. Description of the Drawings

[0055] Figure 1 It is the system flow chart of the present invention; Figure 2 It is the flow chart of multi-time step data decomposition of the present invention; Figure 3 This is the architecture diagram of the multi-scale feature fusion model of the present invention; Figure 4 This is the execution flowchart of the optimization control strategy of the present invention. Detailed implementation manners

[0056] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] Please refer to Figures 1-4 , the present invention provides a technical solution, an optimized operation method for zero-carbon building energy based on reinforcement learning, specifically including the following steps: Step 1. Reasonably collecting and processing multi-source data is the primary step in the optimization control modeling of the multi-energy system of zero-carbon buildings. The data for this experiment comes from the annual data of a certain zero-carbon building from January 1, 2024 to December 31, 2024, covering real-time meteorological data (light intensity, ambient temperature), building energy consumption data (electric / thermal / gas load), and equipment operation status (photovoltaic output, energy storage SOC, heat pump efficiency). When dividing the data set, it is necessary to fully consider the characteristics of the time series to ensure that the data distribution of the training set and the test set can truly reflect the actual energy usage situation and avoid biases caused by seasonal changes or weather characteristics. Specifically, the period from January 1, 2024 to October 31, 2024 (a total of 10 months, 7296 groups of data) is used as the training set; the period from November 1, 2024 to December 31, 2024 (a total of 2 months, 1464 groups of data) is used as the test set.

[0058] In addition, it is also necessary to preprocess the data, including steps such as removing outliers, filling in missing values, and normalization. For outliers, the rule is used for identification and removal; for missing values, linear interpolation is used for filling; the normalization process uses the formula , where is the mean value, and is the standard deviation. To improve the stability and prediction accuracy of the model.

[0059] Reasonably dividing and processing the data set helps to ensure that the model performs well during the training and evaluation processes, providing a solid data foundation for the subsequent optimization control of the energy system. Therefore, scientifically dividing the original data into a training set and a test set is an important prerequisite in the construction of the optimization control model of the multi-energy system of zero-carbon buildings, providing necessary support for the training and verification of the model.

[0060] Step 2: Construct the multi-time-step data hierarchy for the training set data. Through this step, the fluctuation information at different time scales can be effectively retained, enhancing the feature information content of the data set.

[0061] First, aggregate the original 15-minute granular data into 30-minute and 1-hour hierarchies. Specifically as follows: Hierarchy 1 (15 minutes): Retain the original time granularity, with 96 time points in a day. For example, from 0:00 - 0:15, 0:15 - 0:30 on January 1, 2024, etc. Hierarchy 2 (30 minutes): Generate 48 time points through weighted average. The weighted average formula is , where is the data of the t-th time point in Hierarchy 2, and are the data of the corresponding two 15-minute time points in Hierarchy 1, is the time weight coefficient. For example, if = 0.5, then the data from 0:00 - 0:30 on January 1, 2024 is the average of the data from 0:00 - 0:15 and 0:15 - 0:30. Hierarchy 3 (1 hour): Generate 24 time points through aggregation of Hierarchy 2 data. The aggregation formula is , where is the data of the t-th time point in Hierarchy 3, and are the data of the corresponding two 30-minute time points in Hierarchy 2, is the time weight coefficient.

[0062] Secondly, adjust and according to the real-time light intensity: , .

[0063] For example, when the light intensity is strong, appropriately increase the value so that the data of the first 15 minutes accounts for a larger proportion in the aggregation of 30-minute data.

[0064] Step 3: Use the sliding window feature extraction and feature fusion network to process the multi-time-step data. This step can effectively enhance the multi-scale features and provide richer information for subsequent optimal control.

[0065] First, apply a sliding window (window length set to 24 hours) to each hierarchy data for feature extraction. Taking Hierarchy 1 as an example, the sliding window contains 96 15-minute data points, that is, use the 96 data from 0:00 - 23:45 on January 1, 2024 as features. The window starts from the first data in the sequence and slides 1 15-minute data each time until all data is traversed. The sliding window formula is: For the data of Hierarchy k, the feature data of the i-th window in the sliding window is , where n is the window length and t is the starting time point.

[0066] Secondly, construct a three-branch LSTM network to process the data of levels 1-3 respectively, and integrate the features through the attention mechanism. The formula of the attention mechanism is , where is the attention weight of the feature at the k-th level, is the learnable weight matrix, is the output feature of the LSTM network at the k-th level. The fused feature is .

[0067] Step 4: Constructing a multi-objective optimization model is the core step to achieve the coordinated optimization control of the multi-energy system of zero-carbon buildings. Through this step, a balance can be achieved between economy and environmental protection.

[0068] First, determine the objective function. Taking the lowest operating cost and the highest renewable energy consumption rate as the dual objectives, the objective function is .

[0069] Secondly, determine the constraint conditions. Power balance constraint: ; Equipment operation constraint: , ; Energy storage SOC constraint: , .

[0070] Step 5: Generate a rolling optimization strategy for the test set data, and use the same data processing method corresponding to the model parameters obtained by training to achieve the real-time optimization control of the energy system.

[0071] First, input the latest data every 15 minutes to update the energy supply and demand forecast for the next 4 hours (16 time steps). The prediction process is based on the fused features obtained in Step 3 , and the prediction is carried out through a pre-trained prediction model.

[0072] Secondly, use mixed integer linear programming (MILP) to solve the optimal scheduling plan. The goal is to minimize the objective function under the constraint conditions described in Step 4. The optimization problem is expressed as , where is the decision variable (such as the power of each device, the charge and discharge state of the energy storage, etc.), is the objective function, is the equality constraint, is the inequality constraint.

[0073] Finally, send the optimization results to the device controller to adjust the charge and discharge of the energy storage, the operation mode of the heat pump, etc. For example, if it is predicted that the renewable energy generation will be sufficient in the future, the energy storage device can be controlled to charge, that is, when the photovoltaic output > load, the energy storage charges, ; If the load demand is large, the power output of the gas equipment can be increased. That is, when there is a peak load and insufficient photovoltaic power, the gas turbine is started. .

[0074] Step Six: Real-time verification of the prediction results and feedback adjustment are important steps to ensure the optimization control effect of the energy system.

[0075] First, calculate the root mean square error (RMSE) between the predicted value and the measured value in real time. The formula is , where is the predicted value, is the measured value, and N is the number of data points.

[0076] Second, if the error exceeds the threshold, the model retraining is automatically triggered. During the retraining process, update the attention weights in the feature fusion network and optimize the parameters in the model to improve the prediction accuracy and optimization effect of the model.

[0077] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0078] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A collaborative optimization control method for a zero-carbon building energy system, characterized in that, It includes the following steps: Step 1: Collect multi-source data of the building energy system in real time, including meteorological data, energy consumption data, and equipment operation status data; Step 2: Perform multi-time step decomposition on the original 15-minute granularity data to construct a data pyramid containing three levels of 15 minutes, 30 minutes, and 1 hour. The data between levels is weighted and aggregated through dynamic weight coefficients α and γ; Step 3: Use the sliding window technique to extract local features from the data at each level, construct a three-branch LSTM network to process multi-time step features respectively, and achieve multi-scale feature fusion through the attention mechanism; Step 4: Establish an optimization model with the dual objectives of minimizing the operating cost and maximizing the renewable energy consumption rate, and set power balance constraints, equipment operation constraints, energy storage SOC constraints, and carbon emission constraints; Step 5: Use a mixed integer linear programming solver for rolling optimization control, and update the scheduling plan for the next 4 hours every 15 minutes; Step 6: Real-time verification of prediction error and triggering of model retraining, and dynamic update of feature weights and optimization parameters through the RMSE threshold determination mechanism. And optimization parameters.

2. The collaborative optimization control method for a zero-carbon building energy system according to claim 1, characterized in that: The data aggregation between levels in Step 2 uses the following formula: The 15-minute time step is Level 1: The time series is in minutes, corresponding to 96 data points, denoted as ; The 30-minute time step is Level 2: Generated by aggregating level 1 data, with a time series of minutes, corresponding to 48 data points, denoted as ; The 1-hour time step is Level 3: Generated by aggregating level 2 data, with a time series of minutes, corresponding to 24 data points, denoted as ; Each data point at level 2 is obtained by the weighted average of two adjacent data points in level 1: where is the time weight coefficient: When it is an equal-weight average; When focus on the data of the first 15 minutes; Level 3 data point Aggregated from Level 2 data points: wherein is a weight coefficient.

3. A collaborative optimization control method for a zero-carbon building energy system according to claim 1, characterized in that: The formula for local feature extraction in Step 3 is: Where n is the window length and m is the prediction step size.

4. A collaborative optimization control method for a zero-carbon building energy system according to claim 1, characterized in that: The calculation of the attention mechanism uses the following formula: where is the attention weight of the k-th level feature, is the learnable weight matrix, is the output feature of the k-th level LSTM network; The fused feature is .

5. A collaborative optimization control method for a zero-carbon building energy system according to claim 1, characterized in that: The specific content of the dual-objective optimization function in Step 4 includes: Economic objective: Each cost item is calculated based on the time-of-use electricity price, gas price, and carbon price; Environmental protection objective: Among them, the calculation of the renewable energy consumption rate includes a correction term for the curtailment power.

6. The collaborative optimization control method for a zero-carbon building energy system according to claim 1, characterized in that: The specific constraint conditions set in Step 4 include: Power balance equation: ； Equipment operation constraints: ； ； Energy storage SOC dynamic model: ； 。 7. A collaborative optimization control method for a zero-carbon building energy system according to claim 1, characterized in that: The specific implementation method of the rolling optimization strategy in Step 5 is: Each optimization covers the next 16 15-minute time steps, and the optimization variables include the grid interaction power, gas equipment power, energy storage charge and discharge status, and surplus power grid connection power; Use the branch and bound algorithm to solve the MILP problem, and the solution time window constraint is within 3 minutes.

8. A collaborative optimization control method for a zero-carbon building energy system according to claim 1, characterized in that: The feedback mechanism setting in Step 6 includes: Calculate the RMSE value of the future 1-hour prediction result in real time. When the RMSE > 5% of the rated load for three consecutive time steps, trigger model retraining; During the retraining process, use the online gradient descent algorithm to update the attention weight matrix W_k and the LSTM network parameters.

9. A collaborative optimization control method for a zero-carbon building energy system according to claim 2, characterized in that: The specific dynamic weight adjustment strategy includes adjusting according to the real-time light intensity and , and the following formula is adopted: ； 。

Images