A low-energy-consumption green building intelligent operation method based on multi-environment perception and cooperative control

Abstract

The application provides a kind of low energy consumption green building intelligent operation method based on multi-environment perception and collaborative control, relating to building intelligent technology field, method includes: the feature vector of fusion environment state and equipment operation state is constructed, candidate collaborative control strategy is generated using random forest model, and building load and photovoltaic supply are respectively predicted based on double-way prediction logic network, and then the expected net energy consumption corresponding to each control strategy is calculated;Combined with dynamic energy consumption benchmark curve, a constraint optimization mechanism is established to select the optimal control strategy that meets the daily energy consumption limit and convert it into equipment execution instructions to achieve multi-device collaborative regulation;After control execution, the prediction model is corrected online based on actual metering data, and the control sample is positively or negatively labeled according to whether the actual energy consumption is better than the dynamic energy consumption benchmark, which is used for adaptive updating of the random forest model, so as to continuously improve the building energy efficiency and the accuracy of control decision under the premise of ensuring energy consumption constraint.

Description

Technical Field

[0001] This invention relates to the field of building intelligence technology, and in particular to a low-energy green building intelligent operation method based on multi-environment perception and collaborative control. Background Technology

[0002] With the acceleration of urbanization and the continuous improvement of residents' living standards, energy consumption in my country's building sector has maintained rigid growth, indicating huge potential for energy conservation and carbon reduction. Current near-zero energy building technical standards impose stringent requirements on indoor thermal and humidity environments and energy efficiency indicators. Currently, the operation and management of green buildings mainly rely on traditional building automation systems, which collect basic data such as indoor temperature and humidity to automatically control equipment such as air conditioning and lighting to maintain basic indoor comfort.

[0003] Building operation and management is shifting from simply meeting environmental standards to a dual control of environmental quality and energy efficiency. In terms of technological development, the industry is beginning to introduce artificial intelligence algorithms to predict building energy consumption and, in conjunction with the integration of renewable energy sources such as photovoltaics, is attempting to achieve coordinated optimization of power generation, grid, load, and storage. Meanwhile, utilizing multi-dimensional sensors, including those for fine particulate matter concentration and carbon dioxide concentration, to comprehensively perceive indoor environmental quality has become an important feature of high-end green buildings.

[0004] However, existing low-energy building operation methods still face significant technical bottlenecks. Firstly, the control logic exhibits an island effect and lacks multi-variable coordination. Existing solutions often focus on single environmental variables while neglecting the strong coupling relationships between equipment. For example, while activating electric shades reduces air conditioning cooling load, it may lead to insufficient indoor lighting, thus increasing lighting energy consumption. Secondly, there is a disconnect between design specifications and operational data. Theoretical energy consumption indicators calculated during the building design phase cannot dynamically guide actual operation, resulting in actual energy consumption often far exceeding design values. Finally, there is a lack of prediction-based feedforward control mechanisms. Existing systems mostly perform lagging feedback adjustments based on the current state, failing to plan equipment operation strategies in advance based on future weather changes and photovoltaic power generation potential. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a low-energy green building intelligent operation method based on multi-environment perception and collaborative control. This invention solves the problems in the prior art, such as the lack of multi-variable collaboration in control logic, the separation of design indicators and operation data, and the lack of a prediction-based feedforward control mechanism.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for intelligent operation of low-energy green buildings based on multi-environmental perception and collaborative control includes: The monthly energy consumption design index of the target building is obtained, and the monthly energy consumption design index is decomposed into time-series discretization to obtain a dynamic energy consumption benchmark curve that includes daily energy consumption limits. The sensor array is used to collect raw multidimensional environmental data of the target building and the current operating status logs of the controlled equipment; Missing values ​​are imputed and standardized mapping is performed on the original multidimensional environmental data to obtain environmental state feature vectors, and the current running status log is converted into one-hot encoding format to construct a time-series input tensor; The environmental state feature vector is input into a pre-trained random forest classification model to obtain the candidate cooperative control strategy matrix for the next time step. The candidate cooperative control strategy matrix consists of multiple sets of single cooperative control vectors, and each set of single cooperative control vectors is a discrete combination of switch states for the air conditioning, fresh air, lighting and shading systems. A dual-path prediction logic network is constructed based on the time-series input tensor, and the candidate cooperative control strategy matrix is ​​input into the dual-path prediction logic network to calculate the load prediction value and the photovoltaic supply prediction value respectively; the expected net energy consumption value corresponding to a single cooperative control vector is calculated based on the load prediction value and the photovoltaic supply prediction value. A constrained optimization function is constructed based on the difference between the expected net energy consumption value and the dynamic energy consumption benchmark curve; Based on the constraint optimization function, a target single-time coordinated control vector that satisfies the daily energy consumption limit is selected from the candidate coordinated control strategy matrix, and the target single-time coordinated control vector is parsed into a final device execution instruction to drive the controlled device to complete the physical action at the current moment. The actual metering data of the controlled device after responding to the final device execution command is collected, and the dual-path prediction logic network is corrected based on the deviation between the actual metering data and the expected net energy consumption value. Determine whether the actual measurement data is better than the dynamic energy consumption benchmark curve, obtain the determination result, and mark the environmental state feature vector and the target single cooperative control vector as positive and negative sample pairs according to the determination result, so as to update the random forest classification model.

[0007] The present invention discloses the following technical effects: This invention provides a low-energy green building intelligent operation method based on multi-environment perception and collaborative control. The invention introduces a time-series fusion modeling mechanism of multi-environment perception data and equipment status, combines random forest to generate candidate collaborative control strategies, and utilizes a dual-path prediction logic network to decouple load and photovoltaic supply prediction. Under the constraint of a dynamic energy consumption baseline curve, it achieves global optimization of the control strategy, thereby effectively reducing the building's net energy consumption and avoiding the risk of exceeding energy consumption limits. Simultaneously, by introducing a model self-correction and strategy sample incremental evolution mechanism based on actual metering feedback, the control model can continuously and adaptively update with changes in the operating environment, significantly improving the accuracy of energy consumption prediction, the stability of control decisions, and the overall energy-saving level of building operation. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart of a low-energy green building intelligent operation method based on multi-environment perception and collaborative control is provided for an embodiment of the present invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0012] like Figure 1 As shown, this invention provides a low-energy green building intelligent operation method based on multi-environmental perception and collaborative control, comprising: Step 100: Obtain the monthly energy consumption design index of the target building, and perform time-series discretization decomposition on the monthly energy consumption design index to obtain a dynamic energy consumption benchmark curve that includes daily energy consumption limits. Step 200: Use the sensor array to collect raw multidimensional environmental data of the target building and the current operating status logs of the controlled equipment; Step 300: Impute missing values ​​and standardize the original multidimensional environmental data to obtain the environmental state feature vector, and convert the current running status log into a one-hot encoding format to construct a time-series input tensor; Step 400: Input the environmental state feature vector into the pre-trained random forest classification model to obtain the candidate cooperative control strategy matrix for the next time step. The candidate cooperative control strategy matrix consists of multiple sets of single cooperative control vectors. Each set of single cooperative control vectors is a discrete combination of switch states for the air conditioning, fresh air, lighting and shading systems. Step 500: Construct a dual-path prediction logic network based on the time-series input tensor, and input the candidate cooperative control strategy matrix into the dual-path prediction logic network to calculate the load prediction value and the photovoltaic supply prediction value respectively; calculate the expected net energy consumption value corresponding to a single cooperative control vector based on the load prediction value and the photovoltaic supply prediction value. Step 600: Construct a constrained optimization function based on the difference between the expected net energy consumption value and the dynamic energy consumption benchmark curve; Step 700: Based on the constraint optimization function, select the target single-time coordinated control vector that satisfies the daily energy consumption limit from the candidate coordinated control strategy matrix, and parse the target single-time coordinated control vector into the final device execution instruction to drive the controlled device to complete the physical action at the current moment; Step 800: Collect the actual metering data of the controlled device after responding to the final device execution command, and correct the dual-path prediction logic network according to the deviation between the actual metering data and the expected net energy consumption value; Step 900: Determine whether the actual measurement data is better than the dynamic energy consumption benchmark curve, obtain the determination result, and mark the environmental state feature vector and the target single collaborative control vector as positive and negative sample pairs according to the determination result, so as to update the random forest classification model.

[0013] Furthermore, the specific implementation process of step 100 is as follows: In this embodiment, the monthly energy consumption design index corresponding to the unique identifier of the target building in the preset database is first called as the total monthly energy consumption control value, and the start and end dates of the current month are read simultaneously to determine the complete time span. Based on the time span, this embodiment generates a calendar attribute sequence day by day and assigns a date type label to each natural day to distinguish between weekdays and holidays, thereby providing a basic time semantic constraint for subsequent energy consumption allocation. This process only involves calendar rule parsing and database reading, which is a mature data processing method and will not be elaborated further.

[0014] After completing the date type marking, this embodiment retrieves the energy consumption fluctuation pattern corresponding to the building function from a pre-set energy consumption fluctuation feature library according to the building function classification of the target building, and matches the corresponding daily operation weight coefficient in the energy consumption fluctuation feature library according to the date type marking of each day; then, this embodiment normalizes all the obtained daily operation weight coefficients, constructs a normalized time series decomposition vector, so that each weight coefficient satisfies the total amount constraint condition while maintaining the relative energy consumption distribution characteristics, and uses the normalized time series decomposition vector to perform a weighted allocation operation on the monthly total energy consumption control value, thereby obtaining a daily energy consumption discrete limit set corresponding to each natural day in the month.

[0015] After obtaining the daily energy consumption discrete limit set, this embodiment maps each daily energy consumption discrete limit to the same time coordinate system in chronological order. Using the daily time node as the interpolation independent variable and the corresponding energy consumption discrete limit as the interpolation dependent variable, a spline interpolation algorithm is used to smoothly connect adjacent discrete points, thereby generating a dynamic energy consumption benchmark curve that changes continuously in the time dimension. The dynamic energy consumption benchmark curve is used to depict the energy consumption constraint boundary of the target building changing with time in the current month, providing a clear and quantifiable benchmark reference for energy consumption prediction and collaborative control strategy selection in subsequent time periods.

[0016] Furthermore, the specific implementation process of step 200 is as follows: In this embodiment, the multi-protocol aggregation gateway is first initialized and configured, and communication connections are established with various environmental sensors and controlled devices deployed in the target building. During this process, the corresponding communication protocol parsing library is loaded according to the sensor type and device interface protocol, and the device register mapping table corresponding to each controlled device is loaded simultaneously to clarify the mapping relationship between different register addresses and their specific physical meanings. This provides a unified data interpretation basis for subsequent data parsing and status reading. This initialization process is a routine communication configuration, and its specific implementation details will not be elaborated here.

[0017] After initialization, this embodiment periodically sends data request frames to the sensor group through the multi-protocol aggregation gateway according to a pre-set synchronous polling sampling period, and receives raw byte stream response packets returned by each sensor. Subsequently, this embodiment uses the communication protocol parsing library to perform decapsulation and integrity verification operations on the received raw byte stream response packets, extracting the corresponding atomic environmental values ​​one by one, and uniformly attaching a timestamp identifier of the current sampling time to each atomic environmental value to form a raw multidimensional environmental data set that is aligned in the time dimension. The atomic environmental values ​​include at least: temperature, humidity, illuminance, and air quality-related parameters.

[0018] Meanwhile, this embodiment establishes a continuous status monitoring heartbeat link with each controlled device through the multi-protocol aggregation gateway, and reads the status bit sequence in the running and holding register of the controlled device in real time during each sampling period. After obtaining the status bit sequence, this embodiment performs bit-by-bit logical translation of each status bit according to the device register mapping table, converting the underlying binary or numerical status into readable semantic codes representing the device's on / off status, operating gear, and fault status, and records them in chronological order, thereby generating a current running status log that is consistent with the original multidimensional environmental data in the time dimension, providing a reliable data foundation for subsequent time series modeling and collaborative control decisions.

[0019] Furthermore, the specific implementation process of step 300 is as follows: In this embodiment, the original multidimensional environmental data is first traversed point by point according to the time index to detect whether there are any abnormal situations where environmental parameters are missing at each time point. When an abnormal time point is identified, a fixed-length time neighborhood sliding window is constructed with the abnormal time point as the center, and a set of environmental parameter samples that are adjacent in time and have valid values ​​are extracted from the sliding window. The local weighted mean is calculated on the set of valid values, and the result is used to fill the missing data of the corresponding abnormal time point, thereby forming a time-continuous and numerically complete environmental data sequence without introducing additional external data sources.

[0020] After obtaining the complete environmental data sequence, this embodiment further calculates the global mean and global standard deviation of each environmental parameter dimension within the current time span, and performs Z-score normalization mapping on the complete environmental data sequence based on this. This allows environmental parameters with different dimensions and numerical ranges to be uniformly mapped to a comparable standard scale space, thereby constructing a normalized environmental state feature vector that can truly reflect the operating status of the target building environment. This provides stable and numerically consistent input features for subsequent load and supply prediction based on time series models.

[0021] Meanwhile, this embodiment parses the various discrete device state enumeration values ​​recorded in the current operating status log, and establishes a state category index table based on the device state categories that have appeared. Then, according to the index table, the discrete state enumeration value corresponding to each moment is mapped into a high-dimensional sparse binary vector to form a device state one-hot encoding matrix. Subsequently, this embodiment concatenates the normalized environmental state feature vector and the device state one-hot encoding matrix in the feature dimension according to a unified time index to obtain a fused feature matrix. The fused feature matrix is ​​then continuously sliced ​​according to a preset model input step size to construct a time-series input tensor containing time dependencies, which is used for subsequent collaborative control strategy generation and prediction model training.

[0022] Furthermore, the specific implementation process of step 400 is as follows: In this embodiment, the environmental state feature vector obtained in step 300 is first used as a unified input sample and input into a random forest classification model that has been trained offline. At the same time, multiple base decision tree classifiers built in parallel within the random forest classification model are invoked. The environmental state feature vector is synchronously distributed to each base decision tree classifier, so that each base decision tree independently performs discrimination operation according to its own feature subspace partitioning rules, and outputs a local prediction category that matches the current environmental state after traversing to the corresponding leaf node. This local prediction category is used to characterize a more reasonable combination of equipment operating states from the perspective of the base decision tree.

[0023] After obtaining the predicted categories of the local leaf nodes output by all base decision tree classifiers, this embodiment performs soft voting aggregation on the prediction results. By statistically analyzing the frequency of different equipment operating state combination categories in the output results of all base decision trees, a global category probability distribution vector covering all discrete action combinations of air conditioning, fresh air, lighting, and shading systems is constructed. Subsequently, this embodiment sorts the confidence levels in descending order according to the probability values ​​of each category in the global category probability distribution vector, thereby obtaining a strategy index set that reflects the priority of multiple potential collaborative control strategies under the current environmental state.

[0024] After generating the policy index set, this embodiment loads a preset control semantic mapping table and uses the policy index set to perform a key-value reverse lookup operation in the control semantic mapping table to restore each policy index to a specific single-event collaborative control vector. Each set of single-event collaborative control vectors is composed of discrete switching state combinations for air conditioning, fresh air, lighting, and shading systems. Finally, this embodiment sequentially stacks the obtained sets of single-event collaborative control vectors according to the confidence level sorting order to form a candidate collaborative control policy matrix for subsequent prediction, evaluation, and constraint optimization.

[0025] Furthermore, the expression for the random forest classification model is: ; in, This represents the feature vector of the current environmental state. The category of discretized single-stage cooperative control vectors; The number of base decision tree classifiers in the random forest; For the first Each base decision tree outputs a control vector category given an environmental state feature vector. The class probability; This is the global class probability distribution vector obtained after soft voting aggregation of the outputs of all base decision trees; The collaborative control strategy index sequence is obtained by sorting the global category probability distribution vector from high to low.

[0026] Specifically, in this embodiment, the random forest classification model is trained offline based on historical operating data before being put into operation, and is deployed in the control execution unit with fixed model parameters after training. In actual operation, this embodiment uses the environmental state feature vector constructed in step 300 as a single sample input, and simultaneously sends it to multiple base decision tree classifiers set in parallel in the random forest classification model. The base decision tree classifier refers to the classification and discrimination unit independently generated by randomly sampling the sample feature dimension and sample subset during the training phase. Its function is to judge the mapping relationship between the current environmental state and historical control behavior from different feature subspaces.

[0027] When the environmental state feature vector is input to each base decision tree classifier, each base decision tree traverses downwards from the root node according to the feature splitting node threshold stored internally until it reaches the leaf node, and outputs a device operating state combination category and the support probability of that category from the perspective of the base decision tree at the corresponding leaf node. In this embodiment, the device operating state combination categories output by all base decision trees and their corresponding support probabilities are summarized, and the support probabilities of the same device operating state combination category are merged and averaged to obtain a global category probability distribution result covering the entire discrete collaborative control action space. This process is the soft voting aggregation described in this embodiment, which aims to reduce the impact of misjudgment by a single decision tree on the overall decision result and improve the stability and robustness of strategy generation.

[0028] After obtaining the global category probability distribution results, this embodiment sorts them according to the probability magnitude relationship of each device operating state combination category to form a collaborative control strategy index sequence. The device operating state combination category refers to the joint control action composed of the discrete on / off states of the air conditioning, fresh air, lighting, and shading systems at the same time. Subsequently, this embodiment maps the collaborative control strategy index sequence one by one to a specific single collaborative control vector according to the pre-established control semantic mapping table. The single collaborative control vector is used to clearly indicate the specific on / off or gear state that each controlled device should execute at the next time, thereby providing a set of directly executable candidate control strategies for subsequent load prediction, energy consumption assessment, and constraint optimization, ensuring that the output results of the random forest classification model can be correctly understood and executed by the actual control system.

[0029] Furthermore, the specific implementation process of step 500 is as follows: In this embodiment, the dual-path prediction logic network is first initialized using the time-series input tensor constructed in step 300. This dual-path prediction logic network consists of two independent load prediction branches and supply prediction branches. The load prediction branch adopts a long short-term memory network structure to characterize the dynamic evolution characteristics of building load in the time dimension, while the supply prediction branch adopts a multilayer perceptron backpropagation network structure to model the nonlinear mapping relationship between outdoor weather conditions and photovoltaic power output. During initialization, the dual-path prediction logic network loads the corresponding network structure parameters and historical training weights to ensure that stable prediction results can be directly output during the actual inference stage.

[0030] In the load prediction branch, this embodiment performs dimensional alignment processing on each set of single-time cooperative control vectors in the candidate cooperative control strategy matrix with the environmental and equipment state features corresponding to the time-series input tensor at the current time step, and concatenates them along the feature dimensions to form a state-action coupling vector that can simultaneously reflect the current operating state and the control behavior to be executed. Subsequently, this embodiment sequentially inputs the state-action coupling vectors into the Long Short-Term Memory network, and updates the historical state information through the forgetting mechanism, information writing mechanism, and output mechanism inside the loop structure, so that the network outputs the building load prediction result for the next time step corresponding to each set of single-time cooperative control vectors based on the comprehensive influence of historical time-series features and current control actions.

[0031] In the supply forecasting branch, this embodiment extracts meteorological feature slices containing only outdoor temperature, irradiance, and other factors directly related to photovoltaic output from the time-series input tensor. These meteorological feature slices are then input into the multilayer perceptron backpropagation network for forward computation, thereby obtaining the photovoltaic supply forecast value under the current meteorological conditions. This forecast result does not change with the collaborative control strategy. Finally, this embodiment constructs an energy consumption differential calculation logic, performs differential operations on the load forecast value and the photovoltaic supply forecast value at the same time scale, obtains the expected net energy consumption value uniquely corresponding to each single collaborative control vector, and establishes an index association relationship between the two, providing a clear quantitative basis for subsequent energy consumption constraint judgment and strategy.

[0032] Furthermore, the expression for the dual-path prediction logic network is: ; ; ; in, This is the current discrete-time index; The fusion state features are extracted from the temporal input tensor at the current time step; The first candidate in the cooperative control strategy matrix Group single-time cooperative control vector; symbol This indicates a feature dimension concatenation operation; This is the nonlinear mapping function based on the Long Short-Term Memory network in the load prediction branch; In order to execute the first The building load value predicted at the next moment under the condition of a single coordinated control vector; This is a meteorological feature vector containing only outdoor meteorological parameters, obtained by stripping it from the time-series input tensor. For the supply forecasting branch, the mapping function is based on the multilayer perceptron backpropagation network; This is the predicted value of photovoltaic supply at the next moment; In order to be with the first The expected net energy consumption value uniquely corresponds to a single coordinated control vector.

[0033] Specifically, in this embodiment, regarding the specific implementation of the dual-path prediction logic network, this embodiment first extracts the fusion state feature corresponding to the current time from the time-series input tensor at each discrete time step. The fusion state feature refers to the comprehensive feature vector formed by aligning the environmental state feature and the equipment operation state feature in the time dimension, and its function is to fully depict the operating background of the target building at the current moment. Subsequently, in this embodiment, for each single-time collaborative control vector in the candidate collaborative control strategy matrix, the single-time collaborative control vector is concatenated with the fused state features in the feature dimension to form an input feature that couples the state and the control action. This input feature is then fed into the long short-term memory network mapping function in the load prediction branch for processing. The mapping function is used to model the combined impact of historical time-series states and control actions on the building load change at the next moment, thereby outputting the predicted building load value at the next moment under the condition of executing the corresponding single-time collaborative control vector. Meanwhile, this embodiment extracts meteorological feature vectors directly related to photovoltaic power generation, such as outdoor temperature and solar irradiance, from the time-series input tensor. These meteorological feature vectors are then input into the multilayer perceptron backpropagation network mapping function in the supply prediction branch to obtain the photovoltaic supply prediction value for the next moment under the current meteorological conditions. This prediction process does not depend on any specific control strategy. After obtaining the load prediction value and the photovoltaic supply prediction value, this embodiment performs a difference calculation on the two to obtain the expected net energy consumption value corresponding to each set of single-time collaborative control vectors. The expected net energy consumption value is used to characterize the net difference between building energy consumption demand and renewable energy supply within the prediction time step. For example, when the load prediction value is greater than the photovoltaic supply prediction value, the expected net energy consumption value is positive, indicating that energy still needs to be obtained from the external power grid. This provides a clear and quantifiable decision-making basis for subsequent energy consumption constraint judgment and collaborative control strategy.

[0034] Furthermore, the specific implementation process of step 600 is as follows: In this embodiment, to achieve the screening of collaborative control strategies based on energy consumption constraints, at each prediction time, this embodiment first reads the corresponding instantaneous energy consumption threshold from the dynamic energy consumption benchmark curve generated in step 100 according to the current time index. The instantaneous energy consumption threshold refers to the maximum net energy consumption limit that the target building is allowed to consume at that time point, and its function is to provide a clear energy consumption constraint boundary for the feasibility judgment of the collaborative control strategy. Subsequently, for each group of single collaborative control vectors in the candidate collaborative control strategy matrix, this embodiment calls the expected net energy consumption value calculated in step 500 that uniquely corresponds to the single collaborative control vector, and compares the expected net energy consumption value with the instantaneous energy consumption threshold to determine whether the collaborative control strategy meets the energy consumption constraint conditions within the prediction time step.

[0035] When this embodiment determines that the expected net energy consumption value corresponding to a certain single coordinated control vector exceeds the instantaneous energy consumption threshold, it means that after executing the coordinated control strategy, the predicted net energy consumption of the target building will exceed the allowable range defined by the dynamic energy consumption baseline curve. At this time, this embodiment directly assigns an unselectable extremely low strategy evaluation result to the single coordinated control vector, which is used to completely eliminate the strategy in the subsequent optimization process. Conversely, when the expected net energy consumption value is less than or equal to the instantaneous energy consumption threshold, this embodiment calculates the corresponding strategy evaluation score based on the difference between the two. The larger the difference, the more sufficient the energy consumption margin is left under the premise of meeting the energy consumption constraints, and the higher its strategy evaluation score is, thus reflecting the preference for strategies with better energy-saving effects.

[0036] In the aforementioned constraint optimization process, the strategy evaluation score refers to a numerical indicator used to quantify the quality of a single collaborative control vector. It is derived from the relationship between the expected net energy consumption value and the instantaneous energy consumption threshold, and its value is used to establish a comparable priority ranking among multiple candidate collaborative control strategies. For example, when multiple sets of single collaborative control vectors satisfy the energy consumption constraints at the same time, this embodiment prioritizes the strategy with an expected net energy consumption value significantly lower than the instantaneous energy consumption threshold. This ensures that the dynamic energy consumption baseline curve is not exceeded while further reducing the overall energy consumption level of the building, guaranteeing that the constraint optimization function has a clear calculation basis, executability, and stable constraint effect in actual control decisions.

[0037] Specifically, the expression for the constraint optimization function is: ; in, In order to target the The strategy evaluation score obtained from a single collaborative control vector calculation; The first candidate in the cooperative control strategy matrix Group single-time cooperative control vector; The instantaneous energy consumption threshold is obtained by mapping the dynamic energy consumption baseline curve at the current time point.

[0038] Furthermore, the specific implementation process of step 700 is as follows: In this embodiment, for each single-time collaborative control vector in the candidate collaborative control strategy matrix, the expected net energy consumption value for the next moment calculated in step 500 is read, and the expected net energy consumption value is substituted into the constraint optimization logic constructed in step 600 for evaluation one by one. When it is determined that the expected net energy consumption value corresponding to a certain single-time collaborative control vector exceeds the current instantaneous energy consumption threshold mapped by the dynamic energy consumption benchmark curve, this embodiment directly generates a blocking penalty result for the single-time collaborative control vector, which is used to exclude the strategy from the selectable set in the subsequent optimization process. When the expected net energy consumption value is less than or equal to the instantaneous energy consumption threshold, this embodiment generates a corresponding energy-saving fitness score based on the energy consumption margin between the two, thereby forming a quantitative evaluation of the energy consumption compliance and energy-saving effect of each candidate collaborative control strategy at the current moment.

[0039] After evaluating all single-cycle cooperative control vectors, this embodiment aggregates the blocking penalty results or energy-saving fitness scores corresponding to each single-cycle cooperative control vector to construct a strategy evaluation score vector, and performs a maximization optimization sorting process on the strategy evaluation score vector. In this sorting process, this embodiment prioritizes retaining the cooperative control strategy that meets the daily energy consumption limit constraint and has the highest energy-saving fitness score, and uniquely determines the target single-cycle cooperative control vector at the current moment in the candidate cooperative control strategy matrix accordingly, thereby ensuring that the selected control strategy has the best immediate energy-saving effect while meeting the overall energy consumption constraint conditions.

[0040] After determining the target single-time coordinated control vector, this embodiment loads a pre-set device instruction template library and, based on the function codes and control register address information corresponding to each controlled device stored in the template library, parses each discrete switch state combination contained in the target single-time coordinated control vector, converting it into hexadecimal control payloads that conform to the device communication specifications. Subsequently, this embodiment sequentially fills the hexadecimal control payloads into instruction fields that match the corresponding function codes and control register addresses, completes the calculation of cyclic redundancy check codes and the splicing of frame headers and tails, and finally generates device execution instructions that conform to the industrial bus communication standard. These instructions are then sent to the controlled devices through a multi-protocol aggregation gateway to drive each controlled device to complete the corresponding physical action at the current moment.

[0041] Furthermore, the specific implementation process of step 800 is as follows: In this embodiment, after the target single-cycle collaborative control vector is issued and executed, the actual operating metering data of the controlled equipment is collected at the end of the current control cycle. The actual operating metering data includes at least the actual load metering value of the building and the actual power generation metering value of the photovoltaic system. Based on a unified timestamp, this embodiment performs time alignment processing on the actual load metering value and the load prediction value at the corresponding time scale obtained in step 500. At the same time, the actual photovoltaic power generation metering value and the photovoltaic supply prediction value are synchronized to ensure that the prediction results and the actual feedback data are comparable under the same time reference, providing a reliable data foundation for subsequent error calculation and model correction.

[0042] After data alignment is completed, this embodiment constructs a joint loss assessment logic, which incorporates both load prediction error and supply prediction error into a unified correction framework. The load prediction error is used to measure the fitting accuracy of the long short-term memory network to the dynamic changes in building load, while the supply prediction error is used to measure the accuracy of the multilayer sensor network in characterizing changes in photovoltaic output. This embodiment calculates the deviation between the predicted load value and the actual load metering value, as well as the deviation between the predicted photovoltaic supply value and the actual photovoltaic power generation metering value, and forms an error signal for parameter updates based on the deviation, thereby avoiding the continuous accumulation of prediction bias in another prediction branch due to single-path correction.

[0043] During the model calibration phase, this embodiment initiates an error backpropagation mechanism to propagate the timing error signal generated in the load prediction branch back along the time dimension. By gradually correcting the input connection weights and cyclic connection weights of each gate unit in the Long Short-Term Memory network through the backpropagation process over time, the model can more accurately reflect the true impact of the control strategy on load evolution. At the same time, this embodiment propagates the error signal generated in the supply prediction branch back layer by layer along the network hierarchy, calculates the sensitivity of each hidden layer node to the prediction error based on the chain rule, and updates the neuron connection weights in the multilayer perceptron network through gradient descent. This achieves online adaptive calibration of the dual-path prediction logic network without changing the network structure, continuously improving the accuracy of prediction and decision-making in subsequent control cycles.

[0044] Furthermore, the specific implementation process of step 900 is as follows: In this embodiment, after completing the equipment execution and metering data acquisition for the current control cycle, the embodiment first constructs a real energy consumption feedback index based on actual metering data. Specifically, it performs a differential calculation between the actual load metering value and the actual photovoltaic power generation metering value collected in the current cycle to obtain the actual net energy consumption value reflecting the building's actual energy demand level in that cycle. Subsequently, the embodiment reads the energy consumption benchmark limit corresponding to the current time node from the dynamic energy consumption benchmark curve and compares the actual net energy consumption value with the energy consumption benchmark limit to determine whether the building's actual operating energy consumption is better than the pre-set dynamic energy-saving benchmark after executing the target single coordinated control vector, thereby forming a clear and quantifiable basis for energy efficiency judgment.

[0045] When this embodiment determines that the actual net energy consumption value is less than or equal to the energy consumption benchmark limit, it indicates that the current control strategy has achieved an energy-saving effect no higher than the benchmark requirement under real operating conditions. Based on this, this embodiment generates a positive label to characterize that the combination of environmental state and control decision has a positive demonstration significance. Conversely, when the actual net energy consumption value is greater than the energy consumption benchmark limit, it indicates that the control strategy has failed to meet the dynamic energy consumption constraint. This embodiment generates a penalty negative label to indicate that the decision should be given lower priority under the same or similar environmental conditions. Thus, this embodiment obtains the judgment result for supervised learning and ensures that the judgment result comes directly from real operating feedback, avoiding model deviation caused by relying solely on prediction results.

[0046] After generating the judgment result, this embodiment uses the environmental state feature vector corresponding to the current cycle as the feature input, the executed single-time collaborative control vector as the decision output, and the positive label or penalty negative label as the supervision classification label. The three are jointly encapsulated to form incremental evolution sample data. Subsequently, this embodiment writes the incremental evolution sample data into the historical training database of the random forest classification model and triggers the model reconstruction and update process. The feature splitting nodes in each base decision tree classifier are re-evaluated for their classification effect. By recalculating the degree of improvement in class purity brought about by feature splitting, the optimal splitting threshold of each splitting node is dynamically adjusted. This allows the random forest classification model to gradually strengthen the collaborative control strategy that performs well in real energy-saving scenarios and suppress the selection of strategies with poor energy consumption performance, thereby achieving adaptive evolution update for long-term operation.

[0047] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0048] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A low-energy green building intelligent operation method based on multi-environment perception and collaborative control, characterized in that, include: The monthly energy consumption design index of the target building is obtained, and the monthly energy consumption design index is decomposed into time-series discretization to obtain a dynamic energy consumption benchmark curve that includes daily energy consumption limits. The sensor array is used to collect raw multidimensional environmental data of the target building and the current operating status logs of the controlled equipment; Missing values ​​are imputed and standardized mapping is performed on the original multidimensional environmental data to obtain environmental state feature vectors, and the current running status log is converted into one-hot encoding format to construct a time-series input tensor; The environmental state feature vector is input into a pre-trained random forest classification model to obtain the candidate cooperative control strategy matrix for the next time step. The candidate cooperative control strategy matrix consists of multiple sets of single cooperative control vectors, and each set of single cooperative control vectors is a discrete combination of switch states for the air conditioning, fresh air, lighting and shading systems. A dual-path prediction logic network is constructed based on the time-series input tensor, and the candidate cooperative control strategy matrix is ​​input into the dual-path prediction logic network to calculate the load prediction value and the photovoltaic supply prediction value respectively; the expected net energy consumption value corresponding to a single cooperative control vector is calculated based on the load prediction value and the photovoltaic supply prediction value. A constrained optimization function is constructed based on the difference between the expected net energy consumption value and the dynamic energy consumption benchmark curve; Based on the constraint optimization function, a target single-time coordinated control vector that satisfies the daily energy consumption limit is selected from the candidate coordinated control strategy matrix, and the target single-time coordinated control vector is parsed into a final device execution instruction to drive the controlled device to complete the physical action at the current moment. The actual metering data of the controlled device after responding to the final device execution command is collected, and the dual-path prediction logic network is corrected based on the deviation between the actual metering data and the expected net energy consumption value. Determine whether the actual measurement data is better than the dynamic energy consumption benchmark curve, obtain the determination result, and mark the environmental state feature vector and the target single cooperative control vector as positive and negative sample pairs according to the determination result, so as to update the random forest classification model.

2. The intelligent operation method for low-energy green buildings based on multi-environmental perception and collaborative control according to claim 1, characterized in that, The process of obtaining the monthly energy consumption design index of the target building and performing time-series discretization decomposition on the monthly energy consumption design index to obtain a dynamic energy consumption benchmark curve including daily energy consumption limits includes: The system retrieves the monthly total energy consumption control value of the target building from a preset database, extracts the time span characteristics of the current month, and generates a calendar attribute sequence based on the time span characteristics. The calendar attribute sequence includes: date type markers for each day of the current month, and the date type markers include: weekday markers and holiday markers. Based on the building function classification of the target building, a pre-set energy consumption fluctuation feature library is retrieved, and the date type tag is used to match the daily operation weight coefficient corresponding to each day in the energy consumption fluctuation feature library; A normalized time series decomposition vector is constructed based on all the daily operating weight coefficients, and the normalized time series decomposition vector is used to perform weighted allocation calculation on the monthly total energy consumption control value to obtain the daily energy consumption discrete limit set. The daily energy consumption discrete limit set is mapped onto a coordinate system in chronological order, and the discrete points in the daily energy consumption discrete limit set are connected using a spline interpolation algorithm to obtain the dynamic energy consumption benchmark curve.

3. The intelligent operation method for low-energy green buildings based on multi-environmental perception and collaborative control according to claim 1, characterized in that, The process of collecting raw multidimensional environmental data of the target building and current operating status logs of the controlled equipment using a sensor array includes: Initialize the multi-protocol aggregation gateway that connects the sensor group and the controlled device, and load the preset communication protocol parsing library and device register mapping table; According to the preset synchronous polling sampling period, the multi-protocol aggregation gateway broadcasts a data request frame to the sensor group and receives the raw byte stream response packet from the sensor group. The original byte stream response packet is decapsulated and verified using the communication protocol parsing library. Atomic environmental values ​​are extracted and a unified timestamp is added to the atomic environmental values ​​to obtain the original multidimensional environmental data. The atomic environmental values ​​include: temperature, humidity, illuminance and air quality parameters. A status monitoring heartbeat link is established between the multi-protocol aggregation gateway and the controlled device to read the status bit sequence of the controlled device in the running hold register in real time. The status bit sequence is logically translated based on the device register mapping table to convert the status bit sequence into readable semantic codes representing the device's on / off, gear position, and fault status, thereby obtaining the current operating status log.

4. The intelligent operation method for low-energy green buildings based on multi-environmental perception and collaborative control according to claim 1, characterized in that, The process of imputing missing values ​​and standardizing the original multidimensional environmental data to obtain an environmental state feature vector, and converting the current running status log into a one-hot encoded format to construct a time-series input tensor, includes: The time index of the original multidimensional environmental data is traversed to identify abnormal time points where data loss exists, and a time neighborhood sliding window is constructed with the abnormal time points as the center. Extract the effective value set within the time neighborhood sliding window, calculate the local weighted mean of the effective value set, and use the local weighted mean to fill the abnormal time points to obtain a complete environmental data sequence. The global mean and global standard deviation of the complete environmental data sequence are statistically analyzed, and the complete environmental data sequence is subjected to Z-score standardization transformation based on the global mean and global standard deviation to obtain a normalized environmental state feature vector. The discrete state enumeration values ​​in the current running status log are parsed to establish a device status category index table, and the discrete state enumeration values ​​are mapped to high-dimensional sparse binary vectors according to the device status category index table to obtain the device status one-hot encoding matrix. The normalized environmental state feature vector and the device state one-hot encoding matrix are concatenated according to the same time index to obtain a fused feature matrix. The fused feature matrix is ​​then sliced ​​according to a preset model input step size to obtain the temporal input tensor.

5. The intelligent operation method for low-energy green buildings based on multi-environmental perception and collaborative control according to claim 1, characterized in that, The step of inputting the environmental state feature vector into a pre-trained random forest classification model to obtain the candidate cooperative control policy matrix for the next time step includes: The random forest classification model is called with multiple base decision tree classifiers constructed in parallel, and the environmental state feature vector is synchronously distributed to each base decision tree classifier for feature space partitioning, so as to obtain the local leaf node prediction category output by each base decision tree classifier. Soft voting aggregation calculation is performed on all the predicted categories of the local leaf nodes to count the frequency of occurrence of each device operating state combination category, and a global category probability distribution vector covering the entire action space is obtained. The global category probability distribution vectors are sorted by confidence level from highest to lowest probability value to obtain the policy index set; Load the preset control semantic mapping table, and use the policy index set to perform key-value reverse lookup in the control semantic mapping table to obtain multiple sets of single-time cooperative control vectors; The multiple sets of single-time cooperative control vectors are stacked in the order of confidence level to construct the candidate cooperative control strategy matrix; The expression for the random forest classification model is: ； in, This represents the feature vector of the current environmental state. The category of discretized single-stage cooperative control vectors; The number of base decision tree classifiers in the random forest; For the first Each base decision tree outputs the class probability of the control vector category given the feature vector of the environment state. This is the global class probability distribution vector obtained after soft voting aggregation of the outputs of all base decision trees; The collaborative control strategy index sequence is obtained by sorting the global category probability distribution vector from high to low.

6. The intelligent operation method for low-energy green buildings based on multi-environmental perception and collaborative control according to claim 1, characterized in that, A dual-path prediction logic network is constructed based on the time-series input tensor, and the candidate cooperative control strategy matrix is ​​input into the dual-path prediction logic network to calculate the load prediction value and the photovoltaic supply prediction value, respectively. The expected net energy consumption value corresponding to a single cooperative control vector is calculated based on the load prediction value and the photovoltaic supply prediction value, including: Initialize a dual-path prediction logic network containing a load prediction branch and a supply prediction branch, wherein the load prediction branch is configured as a long short-term memory network layer and the supply prediction branch is configured as a multilayer perceptron backpropagation network layer. In the load prediction branch, each single-time cooperative control vector in the candidate cooperative control strategy matrix is ​​dimensionally aligned and concatenated with the current time step feature of the temporal input tensor to obtain a state-action coupling vector. The state-action coupling vector is sequentially input into the long short-term memory network layer. Through iterative calculations using the forget gate, input gate, and output gate, the cell state and hidden layer state are updated to obtain the load prediction value corresponding to each set of single-time cooperative control vectors. In the supply forecasting branch, a meteorological feature slice containing only outdoor meteorological parameters is extracted from the time-series input tensor, and the meteorological feature slice is input into the multilayer sensor backpropagation network layer to obtain a photovoltaic supply forecast value independent of the control strategy. Construct an energy consumption differential calculator, subtract the photovoltaic supply forecast from the load forecast value to obtain the expected net energy consumption value, and establish a unique index association between the expected net energy consumption value and the corresponding single coordinated control vector; The expression for the dual-path predictive logic network is: ； ； ； in, This is the current discrete-time index; The fusion state features are extracted from the temporal input tensor at the current time step; The first candidate in the cooperative control strategy matrix Group single-time cooperative control vector; symbol This indicates a feature dimension concatenation operation; This is the nonlinear mapping function based on the Long Short-Term Memory network in the load prediction branch; In order to execute the first The building load value predicted at the next moment under the condition of a single coordinated control vector; This is a meteorological feature vector containing only outdoor meteorological parameters, obtained by stripping it from the time-series input tensor. For the supply forecasting branch, the mapping function is based on the multilayer perceptron backpropagation network; This is the predicted value of photovoltaic supply at the next moment; In order to be with the first The expected net energy consumption value uniquely corresponds to a single coordinated control vector.

7. A method for intelligent operation of low-energy green buildings based on multi-environmental perception and collaborative control according to claim 6, characterized in that, The expression for the constraint optimization function is: ； in, In order to target the The strategy evaluation score obtained from a single collaborative control vector calculation; The instantaneous energy consumption threshold is obtained by mapping the dynamic energy consumption baseline curve at the current time point.

8. The intelligent operation method for low-energy green buildings based on multi-environment perception and collaborative control according to claim 7, characterized in that, The step of selecting a target single-cycle coordinated control vector that satisfies the daily energy consumption limit from the candidate coordinated control strategy matrix based on the constraint optimization function, and parsing the target single-cycle coordinated control vector into a final device execution command to drive the controlled device to complete the physical action at the current moment, includes: Substitute the expected net energy consumption value corresponding to each group of single-time cooperative control vectors into the constraint optimization function for calculation. If the expected net energy consumption value is greater than the instantaneous energy consumption threshold, then output the blocking penalty value. If the expected net energy consumption value is less than or equal to the instantaneous energy consumption threshold, then output the energy-saving adaptability score. A strategy evaluation score vector is generated based on the blocking penalty value and the energy-saving adaptability score; Maximize the optimization ranking of the policy evaluation score vector to extract a unique target single-round cooperative control vector from the candidate cooperative control policy matrix; Load a pre-set device instruction template library, wherein the device instruction template library stores the function codes and control register addresses of each controlled device; The discrete switch state combinations in the target single-time cooperative control vector are mapped to hexadecimal control payloads, and the hexadecimal control payloads are filled into the fields corresponding to the function code and the control register address. Then, CRC checksum calculation and frame header and tail concatenation are performed to obtain the final device execution instructions that conform to the industrial bus standard.

9. A method for intelligent operation of low-energy green buildings based on multi-environment perception and collaborative control according to claim 6, characterized in that, Collecting actual metering data of the controlled device after responding to the final device's execution command, and correcting the dual-path prediction logic network based on the deviation between the actual metering data and the expected net energy consumption value, including: Read the actual load metering value and the actual photovoltaic power generation metering value for the current time period, and align the actual load metering value with the load forecast value based on the timestamp, and align the actual photovoltaic power generation metering value with the photovoltaic supply forecast value; A joint loss function is constructed that includes a load mean square error term and a supply mean square error term. The joint loss function is used to calculate the time-series deviation gradient between the load forecast value and the actual load measurement value, and the spatial deviation gradient between the photovoltaic supply forecast value and the actual photovoltaic power generation measurement value. The error backpropagation mechanism is initiated to propagate the temporal deviation gradient backward along the time dimension to the long short-term memory network layer, and the input weight matrix and cyclic weight matrix in the gate unit are updated by the backpropagation algorithm over time. The spatial bias gradient is propagated backward along the network hierarchy to the backpropagation network layer of the multilayer perceptron. The error sensitivity of the hidden layer nodes is calculated using the chain rule, and the synaptic weights of the neuron connections are updated using the gradient descent optimizer.

10. A method for intelligent operation of low-energy green buildings based on multi-environmental perception and collaborative control according to claim 9, characterized in that, Determine whether the actual measurement data is better than the dynamic energy consumption benchmark curve, obtain a determination result, and mark the environmental state feature vector and the target single-time cooperative control vector as positive and negative sample pairs according to the determination result, so as to update the random forest classification model, including: The actual net energy consumption value for the current period is obtained by subtracting the actual photovoltaic power generation value from the actual load metering value, and the actual net energy consumption value is compared with the energy consumption benchmark limit value at the current time node in the dynamic energy consumption benchmark curve. If the actual net energy consumption value is less than or equal to the energy consumption benchmark limit, a positive label is generated; if the actual net energy consumption value is greater than the energy consumption benchmark limit, a negative penalty label is generated, and the determination result is obtained. The environmental state feature vector is used as the feature input, the target single-time cooperative control vector is used as the decision output, and the judgment result is used as the supervised classification label. The three are jointly encapsulated to generate incremental evolution sample data. The incremental evolution sample data is injected into the historical training database of the random forest classification model to trigger the model reconstruction mechanism and recalculate the Gini impurity and information gain value of the feature split nodes in all base decision tree classifiers. The optimal splitting threshold of the feature splitting node is adjusted based on the recalculated Gini impurity and the information gain value to complete the adaptive update of the random forest classification model.

Images