A method and system for dynamic control of building thermal balance based on load forecasting

Abstract

This invention relates to the field of intelligent building heat load prediction and thermal balance control technology, and discloses a method and system for dynamic control of intelligent building thermal balance based on load prediction. The method includes: constructing a time-frequency causal dual-flow analysis network to achieve mutual reinforcement of time-frequency features and causal relationships; designing a hierarchical causal discovery algorithm to mine multi-level causal structures; developing a causal-enhanced time-frequency representation learning method to fuse time-frequency and causal information; establishing an intervention decision-making framework based on counterfactual analysis to evaluate the intervention effect; implementing an anomaly pattern self-evolution identification system to continuously learn new anomaly patterns; and developing an interpretable anomaly diagnosis mechanism to provide clear diagnostic reports. This invention improves the accuracy of heat load prediction, enhances anomaly early warning capabilities, increases intervention efficiency, enhances system interpretability, and optimizes energy utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent building heat load prediction and heat balance control technology, and more specifically, to an intelligent building heat balance dynamic control method and system based on load prediction. Background Technology

[0002] With the rapid development of smart buildings, building heat load prediction and heat balance control technologies have received widespread attention. Current heat load prediction mainly relies on statistical analysis and machine learning methods. While these methods can achieve prediction to a certain extent, they lack an understanding of the system's inherent causal mechanisms and are therefore ill-equipped to handle complex and ever-changing real-world operating environments.

[0003] Traditional thermal balance control systems typically employ rule-based control strategies. This approach relies heavily on expert experience and struggles to adapt to dynamically changing environmental conditions. Furthermore, existing anomaly detection and diagnosis methods primarily focus on superficial characteristics, failing to effectively trace the root causes of anomalies. This leads to system response delays and negatively impacts control effectiveness.

[0004] Furthermore, existing technologies generally suffer from insufficient interpretability; the system's decision-making process remains a black box for users, which not only reduces user trust but also increases the difficulty of system maintenance and optimization. The system lacks adaptive learning capabilities, making it difficult to cope with emerging abnormal patterns and limiting its effectiveness in practical applications. Summary of the Invention

[0005] This invention provides a method and system for dynamic control of building thermal balance based on load forecasting, which solves the technical problems in related technologies such as lack of understanding of deep causal mechanisms, insufficient prediction accuracy, lagging anomaly detection, poor interpretability and limited adaptive capability.

[0006] This invention provides a method for dynamic control of heat balance in smart buildings based on load forecasting, comprising the following steps:

[0007] S1: Construct a time-frequency-causal dual-stream analysis network, comprising: a time-frequency analysis branch, consisting of multiple convolutional and pooling layers, employing one-dimensional convolution to extract sequence features along the time axis and two-dimensional convolution to extract local patterns on the time-frequency plane; a causal analysis branch, consisting of a graph convolutional network, with the input being the adjacency matrix and node feature matrix of a preliminary causal graph; and an attention fusion module, consisting of a multi-head self-attention mechanism and a cross-attention mechanism; simultaneously performing time-frequency characteristic analysis and causal relationship analysis on the input heat load time-series data, outputting an enhanced heat load representation after dual-stream analysis;

[0008] S2: Based on the enhanced heat load representation of S1 output, the fit between the causal structure and the data is evaluated by the Bayesian scoring function. A greedy search strategy is applied to explore the possible causal structure space. Combined with the temporal constraints of the time series, direct causal relationships, indirect causal chains and common cause structures are identified to generate a complete causal graph of the heat load system.

[0009] S3: Based on the causal graph generated by S2 and the time-frequency features of S1, a causal-enhanced time-frequency representation learning model is constructed. By weighting the time-frequency domain weights with the outputs of multiple causal time-frequency feature extractors, including short-term fluctuation feature extractors, daily cycle feature extractors, long-term trend feature extractors and abnormal pattern feature extractors, the causal-enhanced time-frequency representation is output.

[0010] S4: Based on the causal enhanced time-frequency representation of S3, construct a structural causal model of the heat load system, remove the incoming edges of the intervened variable through do-calculus operation, calculate the conditional probability distribution of other variables in the system according to the modified causal graph, evaluate the effects of different intervention strategies and select the optimal intervention scheme;

[0011] S5: Based on the causal enhanced time-frequency representation of S3 and the intervention scheme of S4, an integrated learning framework including decision tree, support vector machine and neural network is constructed. The online update of model parameters is realized through incremental learning algorithm, and new abnormal patterns are identified by density clustering algorithm. The abnormal identification results and updated abnormal pattern library are output.

[0012] S6: Based on the anomaly identification results of S5 and the cause-effect graph of S2, by identifying all possible paths from potential cause nodes to anomaly manifestation nodes, calculating the weight and path contribution of each path, and generating a multi-level anomaly diagnosis report including technical, management and operational layers.

[0013] Furthermore, the time-frequency analysis branch in step S1 specifically includes:

[0014] Select Morlet wavelet or Meyer wavelet as wavelet basis function, perform continuous wavelet transform on indoor temperature, supply and return water temperature, valve opening degree and outdoor meteorological parameters in the input heat load time series data to obtain the time spectrum of each variable;

[0015] Calculate the coherence coefficient between spectra for different variables, and identify potential associations when the coherence coefficient is greater than a preset threshold;

[0016] The causal analysis branch determines the conditional independence between variables by calculating the partial correlation coefficient of time series data, applies the PC algorithm to infer the causal direction between variables based on the conditional independence test, and generates the adjacency matrix of the preliminary causal graph.

[0017] The attention fusion module calculates the correlation weights between time-frequency features and causal features, and achieves mutual enhancement of features through weighted combination.

[0018] Furthermore, the Bayesian scoring function in step S2 adopts either BIC scoring or BDeu scoring, and the greedy search strategy includes:

[0019] Initialize an empty cause-effect graph, add or delete edges step by step, and calculate the score value after each modification;

[0020] The search stops when the score no longer increases, thus obtaining a locally optimal causal structure.

[0021] The timing constraints include: the timestamp of the cause variable must be earlier than or equal to the timestamp of the result variable;

[0022] The direct causal relationship is defined as the existence of a direct causal edge connection between variables, the indirect causal chain is defined as a causal path formed through intermediate variables, and the common cause structure is defined as a structure in which multiple variables are affected by the same upstream variable.

[0023] Furthermore, the causal time-frequency feature extractor in step S3 specifically includes:

[0024] Short-term fluctuation feature extractor: It adopts a high-pass filter combined with a one-dimensional convolutional neural network, with a convolutional kernel size of 5-15 minutes time span, and includes 3 convolutional layers and 2 pooling layers;

[0025] Daily Periodic Feature Extractor: Employs a periodic convolution structure, with the convolution kernel reflecting the 24-hour periodicity, and extracts periodic patterns through Fourier transform;

[0026] Long-term trend feature extractor: Combining wavelet transform and LSTM network, the LSTM has 128 units and contains a 2-layer stacked structure;

[0027] Anomaly pattern feature extractor: It adopts an autoencoder structure, with the encoder and decoder each containing 3 fully connected layers, and the bottleneck layer has a dimension of 32;

[0028] The weighted combination is achieved through an adaptive gating mechanism, which dynamically adjusts the weight coefficients of each feature extractor according to the current heat load status.

[0029] Furthermore, the structural causal model in step S4 is constructed in the following manner:

[0030] Based on the cause-effect graph, a structural equation is defined between variables, where each endogenous variable is represented as a function of its parent node and the noise term;

[0031] The parameters of the structural equation are estimated using historical data, employing either the least squares method or maximum likelihood estimation.

[0032] The do-calculus operation specifically involves: removing all incoming edges of the variable to be intervened from the causal graph, setting the variable as the intervention value, and calculating the posterior distribution of other variables through probabilistic inference;

[0033] The evaluation system includes energy efficiency indicators, indoor temperature comfort indicators, and system stability indicators, which are weighted and summed to obtain a comprehensive score.

[0034] Furthermore, the specific implementation of the do-calculus operation includes:

[0035] Construct a causal graph G before intervention and a causal graph G' after intervention, where G' is obtained by removing all edges from G that point to the intervened variable;

[0036] Let the variable being intervened be X, and the intervention value be x. For any variable Y, the probability distribution P(Y|do(X=x)) after intervention is calculated as follows:

[0037] If Y is an ancestor node of X, then P(Y|do(X=x))=P(Y);

[0038] If Y is a descendant node of X, then the conditional probability is calculated using the d-separation criterion and the chain rule.

[0039] Monte Carlo simulation was used to generate confidence intervals for estimating the intervention effect using 1000 samples.

[0040] Furthermore, the ensemble learning framework in step S5 specifically includes:

[0041] The decision tree classifier uses the C4.5 algorithm to handle anomalous features with well-defined rules;

[0042] Support vector machines use radial basis kernel functions to handle anomaly categories with clear boundaries;

[0043] The neural network uses a 3-layer fully connected structure to handle complex nonlinear features;

[0044] The outputs of each classifier are integrated through a weighted voting mechanism, with the weights determined by the performance of the validation set.

[0045] The incremental learning algorithm employs the incremental decision tree algorithm for decision trees, the online SVM algorithm for support vector machines, and gradient accumulation and model fine-tuning for neural networks.

[0046] The density clustering algorithm uses DBSCAN to automatically determine the number of clusters and identify new anomaly categories.

[0047] Furthermore, the path contribution calculation method in step S6 includes:

[0048] Identify all paths from potential cause nodes to anomalous behavior nodes using depth-first search or breadth-first search algorithms;

[0049] Calculate the weight value for each edge, which is equal to the product of the causal strength of that edge and the temporal correlation coefficient;

[0050] The path contribution of each path is the product of the weights of all edges on the path.

[0051] The paths are sorted in descending order of their contribution, and the paths with a contribution greater than a threshold are selected as the main anomaly propagation paths.

[0052] The multi-level diagnostic report is generated through template matching. The technical layer contains detailed parameters and causal paths, the management layer contains an anomaly overview and impact analysis, and the operational layer contains specific maintenance recommendations.

[0053] Furthermore, the method for calculating the causal strength includes:

[0054] For linear causal relationships, the causal strength is equal to the path coefficient in the structural equation;

[0055] For nonlinear causal relationships, the causal strength is determined by calculating the average causal effect or the conditional average causal effect.

[0056] The time-series correlation coefficient is obtained by calculating the maximum value of the cross-correlation function under different time delays;

[0057] The path contribution threshold is set to 0.5 times the average contribution of all paths to ensure that only the main anomaly propagation paths are retained.

[0058] This invention also provides a smart building thermal balance dynamic control system based on load forecasting, comprising:

[0059] The time-frequency-causal dual-stream analysis module includes: a time-frequency analysis submodule, consisting of multiple convolutional and pooling layers, containing one-dimensional and two-dimensional convolutional units; a causal analysis submodule, consisting of a graph convolutional network, containing an adjacency matrix input unit and a node feature processing unit; and an attention fusion submodule, consisting of multi-head self-attention units and cross-attention units.

[0060] The hierarchical causal discovery module includes: a Bayesian scoring unit for calculating BIC or BDeu scores; a search strategy unit for executing a greedy search algorithm; a temporal constraint unit for ensuring the temporal order of causal relationships; and a causal relationship classification unit for identifying direct causal chains, indirect causal chains, and common cause structures.

[0061] The causal enhanced time-frequency representation learning module includes: a short-term fluctuation feature extractor, containing a high-pass filter and a 3-layer convolutional network; a daily periodic feature extractor, containing a periodic convolutional layer; a long-term trend feature extractor, containing a wavelet transform unit and a 2-layer LSTM network; an abnormal pattern feature extractor, containing an autoencoder with a 3-layer encoder and a 3-layer decoder; and an adaptive gating unit for dynamically adjusting the weights of each feature extractor.

[0062] The counterfactual analysis module includes: a structural causal model building unit for defining structural equations between variables; a do-calculus operation unit for removing incoming edges from the causal graph and calculating the probability distribution after intervention; an evaluation index calculation unit for calculating energy efficiency, comfort, and stability indicators; and an intervention strategy selection unit for selecting the optimal solution based on a comprehensive score.

[0063] The anomaly pattern self-evolution identification module includes: an ensemble classifier containing decision trees, support vector machines, and neural networks; an incremental learning unit for online updating of model parameters; a density clustering unit that uses the DBSCAN algorithm to identify new anomaly categories; and a knowledge base management unit for storing and updating anomaly patterns.

[0064] The anomaly diagnosis module includes: a path search unit for identifying causal propagation paths; a contribution calculation unit for calculating path weights and path contributions; a multi-level report generation unit for generating diagnostic reports at the technical, management, and operational levels; and a diagnostic visualization unit for displaying causal paths and anomaly characteristics.

[0065] The beneficial effects of this invention are as follows:

[0066] This invention improves the accuracy of heat load prediction by employing an innovative time-frequency causal dual-flow analysis network. It not only captures the time-frequency characteristics of data but also provides a deeper understanding of the system's inherent causal mechanisms, making the prediction results more reliable and stable.

[0067] By employing a hierarchical causal discovery algorithm and causal-enhanced time-frequency representation learning, a deeper understanding and more precise control of the heat load system are achieved. The system can automatically discover multi-level causal relationships and effectively integrate this information into the time-frequency representation, thereby improving the system's predictive and control capabilities.

[0068] Based on a counterfactual analysis framework and an anomaly pattern self-evolutionary identification system, this invention improves the system's anomaly early warning and handling capabilities. The system can detect potential problems in advance and continuously improve its prevention and handling capabilities by constantly learning and adapting to new anomaly patterns.

[0069] An explainable anomaly diagnosis mechanism enhances the interpretability and credibility of the system. Through clear causal path analysis and multi-level diagnostic interpretation, the system's decision-making process becomes more transparent, facilitating understanding by managers and execution by operators, while simultaneously improving energy efficiency and indoor environmental comfort. Attached Figure Description

[0070] Figure 1 This is a flowchart of a smart building thermal balance dynamic control method based on load forecasting according to the present invention;

[0071] Figure 2 This is a bar chart comparing the accuracy of heat load prediction in this invention;

[0072] Figure 3 This is a bar chart comparing the advance warning time of the anomaly detection method according to the present invention;

[0073] Figure 4 This is a radar chart showing the efficiency of the intervention strategy of this invention;

[0074] Figure 5 This is a bar chart comparing the system interpretability evaluation of the present invention;

[0075] Figure 6 This is a line graph showing the change in the anomaly identification accuracy of the present invention over time;

[0076] Figure 7 This is a bar chart showing the energy consumption optimization effects of the present invention on different building types. Detailed Implementation

[0077] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0078] At least one embodiment of the present invention discloses a method for dynamic control of heat balance in smart buildings based on load forecasting, such as... Figure 1 As shown, it includes the following steps:

[0079] Step 1: Construct a time-frequency causal dual-flow analysis network, including a time-frequency domain analysis flow and a causal relationship analysis flow, to simultaneously process the time-frequency characteristics and causal relationships of heat load time-series data, and output preliminary time-frequency characteristics and causal relationship graphs;

[0080] Specifically, the following steps are included:

[0081] Step 1.1: Input heat load time series data, including multi-source data such as indoor temperature, supply and return water temperature, valve opening, and outdoor meteorological parameters;

[0082] Data preprocessing includes missing value imputation, outlier filtering, and time alignment.

[0083] Step 1.2: Construct a time-frequency domain analysis stream and use an improved continuous wavelet transform algorithm to process the heat load time series data to obtain time-frequency representations with different time scales and frequency resolutions.

[0084] The specific steps are as follows:

[0085] Choose a wavelet basis function that suits the characteristics of the heat load data, such as Morlet wavelet or Meyer wavelet;

[0086] Perform a continuous wavelet transform on each heat load-related variable to obtain its time-frequency spectrum;

[0087] Calculate the coherence between spectra for different variables to identify potential associations.

[0088] Step 1.3: Construct causal relationship analysis flow. Use PC algorithm and conditional independence test to construct preliminary causal graph structure to represent the causal relationships in the heat load system.

[0089] The specific steps are as follows:

[0090] Calculate the conditional independence relationship between variables based on time series data;

[0091] Applying PC algorithms to infer the causal direction between variables;

[0092] The initial cause-effect graph is revised and verified by incorporating domain knowledge.

[0093] Step 1.4: Implement the attention fusion mechanism to enable the two analysis streams to mutually enhance each other;

[0094] Time-frequency analysis streams provide temporal features and frequency information for causal discovery, assisting in the identification of causal edges;

[0095] Cause-effect graph structures provide a focus for time-frequency analysis and optimize time-frequency feature extraction parameters;

[0096] Cross-attention mechanism enables information exchange and feature enhancement between two streams.

[0097] The aforementioned two-stream analysis network structure includes:

[0098] Time-frequency analysis branch: Composed of multiple convolutional layers and pooling layers, used to extract time-frequency features at different scales.

[0099] This branch uses a combination of one-dimensional and two-dimensional convolutions. One-dimensional convolutions extract sequence features along the time axis, while two-dimensional convolutions extract local patterns on the time-frequency plane.

[0100] The network has a depth of 5 layers. Each convolutional layer is followed by a batch normalization and linear correction unit (ReLU) activation function to ensure the non-linear expressive power of feature extraction.

[0101] Causal analysis branch: Composed of graph convolutional networks (GCNs), used to handle causal relationships between variables.

[0102] GCN consists of three graph convolutional layers. Each convolutional operation aggregates information from neighboring nodes to update node representations. The input consists of the adjacency matrix and node feature matrix of the initial causal graph.

[0103] Attention fusion module: Composed of multi-head self-attention mechanism and cross-attention mechanism, it enables mutual enhancement of information from the two branches. The attention module calculates the correlation weights between different features, focusing on the most relevant feature combinations.

[0104] In some embodiments, the graph convolutional network can be replaced with a graph attention network (GAT) to enhance the ability to focus on important causal relationships.

[0105] Furthermore, the attention fusion module can employ different attention mechanisms, such as, optionally, a self-attention mechanism based on a Transformer architecture.

[0106] Output: Enhanced heat load representation after dual-flow analysis, including time-frequency characteristics and causal structure information, providing a basis for subsequent processing.

[0107] like Figure 2 As shown, the accuracy of heat load prediction using different methods at different time scales is compared. It can be seen that the causal time-frequency fusion technology proposed in this invention significantly outperforms traditional correlation analysis and single time-frequency analysis methods in short-term (1 hour), medium-term (24 hours), and long-term (7 days) predictions, with an average improvement in prediction accuracy of about 45%, verifying the "significantly improved prediction accuracy" mentioned in the patented technology effect.

[0108] Step 2: Based on time-frequency features and causal relationship graphs, establish a hierarchical causal discovery algorithm to discover multi-level causal relationships, including direct causal relationships, indirect causal chains, and common cause structures, and output the complete causal structure;

[0109] Specifically, the following steps are included:

[0110] Step 2.1: Construct a scoring-based causal structure learning method, optimized for the characteristics of time-series heat load data: - Define a Bayesian scoring function to evaluate the fit between the causal structure and the data - Apply a greedy search strategy to efficiently explore the possible causal structure space - Combine the temporal constraints of the time series to ensure that the temporal order of the causal relationship is logically reasonable.

[0111] Step 2.2: Implement multi-level causal relationship mining and distinguish different types of causal relationships;

[0112] Identify direct causal relationships, such as outdoor temperature directly affecting water supply temperature;

[0113] Indirect causal chains were discovered, such as outdoor temperature indirectly affecting indoor temperature through heat storage in the walls.

[0114] Detect common cause structures, such as when the same control strategy simultaneously affects multiple system parameters.

[0115] Step 2.3: Correct and verify the cause-effect graph by incorporating domain knowledge;

[0116] Incorporating the fundamental principles of building thermodynamics as prior knowledge;

[0117] We apply a time-lapse-based causal verification method to test causal hypotheses.

[0118] Cross-validation was used to evaluate the reliability and stability of the causal discovery results.

[0119] In some embodiments, in addition to Bayesian scoring functions, information theory-based scoring functions, such as the Minimum Description Length (MDL) criterion or the mutual information criterion, may also be used.

[0120] In addition, for scenarios with limited computing resources, constrained basis algorithms, such as PC algorithm variants based on local learning, can be selected to reduce computational complexity.

[0121] Output: A complete cause-effect graph of the heat load system, containing multi-level causal relationships between variables, and labeling the strength and confidence of the relationships.

[0122] like Figure 3 As shown, the comparison of early warning lead times for various anomaly detection methods is illustrated. It can be seen that the causal enhancement anomaly detection technology proposed in this invention provides significantly longer early warning lead times than traditional anomaly detection methods in four scenarios: equipment malfunction, abnormal energy consumption fluctuations, control strategy anomalies, and environmental impact anomalies. The average lead time is increased by approximately 300%, verifying the "significantly enhanced anomaly early warning capability" mentioned in the patented technology's effects.

[0123] Step 3: Utilize causal structure and time-frequency features to achieve causal-enhanced time-frequency representation learning, integrate causal information into time-frequency analysis, and output causal-enhanced time-frequency representation;

[0124] Specifically, the following steps are included:

[0125] Step 3.1: Establish a causal reinforced time-frequency representation learning (CETFR) mathematical model;

[0126] The CETFR expression is defined as a comprehensive computation method that combines time-frequency domain weights with the outputs of multiple causal time-frequency feature extractors.

[0127] Specifically:

[0128] Obtain the time-frequency domain weight value of the current data at a specific time and frequency point. This weight value reflects the importance of different time and frequency points.

[0129] Calculate the output values ​​of multiple causal time-frequency feature extractors under the conditions of current data, current time point, current frequency point, and system causal graph.

[0130] The output value of each feature extractor is multiplied by its corresponding weight coefficient, which is determined through model training and reflects the importance of each feature extractor.

[0131] The summation of all weighted feature extractor outputs is then multiplied by the time-frequency domain weights to obtain the final causal-enhanced time-frequency representation.

[0132] This representation method can simultaneously consider the time-frequency characteristics and causal structure information of the data, and achieve comprehensive capture of the multifaceted characteristics of the data by weighted combination of the outputs of multiple specialized feature extractors.

[0133] The specific implementation of the causal time-frequency feature extractor is as follows:

[0134] Applying specific time-frequency domain transformations (such as short-time Fourier transform or wavelet transform) to the input data yields a preliminary time-frequency representation;

[0135] The focus weights are calculated based on the strength of relationships in the causal graph, with particular attention paid to variables and time-frequency regions with strong causal relationships.

[0136] The enhanced feature representation is obtained by multiplying the initial time-frequency representation with the attention weights and processing it through a nonlinear activation function.

[0137] Different feature extractors correspond to different time-frequency scales and feature extraction focuses, together forming a multi-scale feature extraction system.

[0138] Step 3.2: Design multiple causal time-frequency feature extractors to target different characteristics of heat load data;

[0139] Short-term fluctuation feature extractor, focusing on heat load changes at the minute to hour level;

[0140] Daily cycle feature extractor, focusing on load patterns within a 24-hour cycle;

[0141] Long-term trend feature extractor, focusing on heat load changes from daily to monthly levels;

[0142] Anomaly pattern feature extractor focuses on identifying signal features that deviate from normal patterns.

[0143] Step 3.3: Optimize time-frequency analysis parameters based on causal structure;

[0144] The key nodes and edges in the causal graph are used to guide the region of interest in time-frequency analysis;

[0145] The time window and frequency resolution of the wavelet transform are dynamically adjusted based on the strength of the causal relationship.

[0146] Choose the most suitable wavelet basis functions and parameters for different types of causal relationships.

[0147] Step 3.4: Achieve adaptive fusion of causal time-frequency features;

[0148] An adaptive weighting calculation mechanism is designed to dynamically adjust the weights of each feature extractor based on the current heat load status.

[0149] The importance of different time points and features is weighted using an attention mechanism;

[0150] A multi-level fusion strategy is constructed, which gradually integrates low-level time-frequency features to high-level semantic features.

[0151] The specific implementation of the causal time-frequency feature extractor includes:

[0152] Short-term fluctuation feature extractor: It adopts a high-pass filter combined with a one-dimensional convolutional neural network structure, with the convolutional kernel size set to a time span of 5 to 15 minutes, and contains 3 convolutional layers and 2 pooling layers, specifically to capture rapidly changing heat load fluctuation features.

[0153] The input is normalized time series data, and the output is a feature vector representing short-term fluctuation characteristics.

[0154] Daily Periodic Feature Extractor: Employs a periodic convolution structure with a convolution kernel design that reflects the 24-hour periodicity, and extracts periodic patterns through Fourier transform and harmonic analysis.

[0155] The extractor includes periodic fill layers, periodic convolutional layers, and global pooling layers, with a particular focus on intraday load variation patterns.

[0156] Long-term trend feature extractor: Combining wavelet transform and long short-term memory (LSTM) network, trend features are extracted after downsampling long-term series data.

[0157] The LSTM has 128 cells and contains a 2-layer stacked structure. It captures slow changes from a day to a month scale by setting the time step.

[0158] Anomaly pattern feature extractor: Employs an autoencoder structure, consisting of an encoder and a decoder. The encoder compresses the input data into a low-dimensional latent space, while the decoder attempts to reconstruct the original data.

[0159] The reconstruction error is used to identify abnormal patterns. The encoder consists of 3 fully connected layers, with a bottleneck layer dimension of 32. The decoder mirrors the encoder structure.

[0160] The adaptive fusion network employs a gating mechanism, which includes:

[0161] Multiple feature gating units, each corresponding to a feature extractor;

[0162] The gating unit generates weight values ​​between 0 and 1 through a fully connected layer and a sigmoid activation function;

[0163] Context-aware gating mechanism dynamically adjusts weights based on current heat load status and historical data;

[0164] The residual connection structure ensures that important information is not lost during the fusion process.

[0165] In some embodiments, sequence models different from LSTM, such as gated recurrent units (GRUs) or bidirectional LSTMs, can be used to enhance the ability to extract long-term trend features.

[0166] In addition, the parameter configuration of each feature extractor can be adjusted according to the specific application scenario. For example, for scenarios with more significant high-frequency fluctuations, the weight of the short-term fluctuation feature extractor in the fusion process can be increased.

[0167] Output: A causal-enhanced time-frequency representation of heat load, containing multi-scale time-frequency features guided by causal structure, providing high-quality input for anomaly detection and prediction.

[0168] like Figure 4As shown, the efficiency of traditional intervention strategies and the counterfactual analysis-based intervention strategy proposed in this patent is compared across five dimensions: response time, energy consumption, stability, energy-saving effect, and user comfort. It can be seen that the counterfactual analysis intervention is significantly superior to the traditional intervention strategy in all dimensions, with an overall efficiency improvement of approximately 60%, verifying the "significantly improved intervention efficiency" mentioned in the patent's technical effect description.

[0169] Step 4: Based on the time-frequency representation of causal reinforcement, construct a counterfactual analysis framework, evaluate the effectiveness of the intervention strategy through do-calculus operation, and output the optimal intervention plan;

[0170] Specifically, the following steps are included:

[0171] Step 4.1: Construct a structural causal model (SCM) for the heat load system.

[0172] Define the functional relationships between variables based on the discovered causal graphs;

[0173] Estimate the parameters of each functional relationship to represent the strength of the causal mechanism;

[0174] Model validation ensures that the SCM can accurately simulate system behavior.

[0175] Step 4.2: Simulate a randomized intervention and evaluate the intervention effect;

[0176] Define the set of possible intervention actions, such as adjusting the water supply temperature, changing the valve opening, etc.

[0177] Perform do-calculus operations on each potential intervention point to simulate the system state after intervention;

[0178] Generate predictions of system behavior under various intervention scenarios to form counterfactual samples.

[0179] do-calculus is an interventional calculus method, and its specific implementation is as follows:

[0180] Removing all incoming edges of the intervened variable in the causal-effect diagram (i.e., the causal influence of other variables on this variable) indicates that the variable is no longer affected by its original factors, but has been artificially set to a specific value.

[0181] Calculate the conditional probability distribution of other variables in the system based on the modified cause-effect graph;

[0182] Generate predictions of the overall state of the system after intervention using probabilistic reasoning or Monte Carlo simulation methods.

[0183] This method can distinguish between correlation and causation, and accurately assess the actual effect of intervention actions.

[0184] Step 4.3: Develop an indicator system for evaluating the intervention effect;

[0185] Define energy efficiency indicators and assess the impact of interventions on energy consumption;

[0186] Design comfort indicators and assess the impact of interventions on the indoor environment;

[0187] Construct stability indicators to assess the impact of interventions on the system's equilibrium state;

[0188] Calculate the overall score and balance the goals across multiple dimensions.

[0189] Step 4.4: Establish the optimal intervention strategy selection algorithm;

[0190] Different intervention strategies were ranked based on evaluation indicators;

[0191] Considering implementation costs and technical feasibility, filter out unsuitable intervention options;

[0192] Select the intervention strategy with the highest overall benefits and generate a specific implementation plan.

[0193] In some embodiments, Markov decision processes or partially observable Markov decision processes can be used to model interventional decision problems in order to address situations where the system state is partially observable.

[0194] In addition, reinforcement learning methods, such as model-based reinforcement learning or Monte Carlo tree search, can be combined to optimize the intervention strategy selection process.

[0195] Output: The optimal intervention strategy for the current heat load state, including the intervention point, intervention parameters, and expected results.

[0196] like Figure 5 As shown, the interpretability of the causal path analysis system proposed in this paper is compared with that of the traditional system from four dimensions: technical personnel understanding, management satisfaction, user trust, and decision-making process transparency. It can be seen that the causal path analysis system significantly outperforms the traditional system in all dimensions, especially in terms of decision-making process transparency, thus verifying the "comprehensive enhancement of system interpretability" mentioned in the patent technology's effect description.

[0197] Step 5: Based on the time-frequency representation of causal reinforcement and the intervention scheme, realize the abnormal pattern self-evolution recognition system, continuously learn new abnormal patterns through multi-module collaboration, and output the abnormal recognition results;

[0198] Specifically, the following steps are included:

[0199] Step 5.1: Establish a basic exception pattern library;

[0200] Collect and label common abnormal heat load patterns, such as equipment failure and abnormal energy consumption;

[0201] Extract the time-frequency and causal features of each anomaly pattern to form a feature template;

[0202] Construct an initial anomaly classifier to identify known anomalies based on template matching.

[0203] Step 5.2: Implement the online anomaly detection algorithm;

[0204] Continuously monitor the causal enhancement time-frequency representation from step 3;

[0205] Calculate the deviation of the current representation from the normal pattern and quantify the degree of anomaly;

[0206] A multi-level threshold strategy is applied to achieve different levels of anomaly warning.

[0207] Step 5.3: Construct a new anomaly pattern learning mechanism;

[0208] When an unknown abnormal pattern is detected, its feature representation is automatically saved;

[0209] When similar unknown anomalies occur repeatedly, clustering and pattern summarization are performed;

[0210] New abnormal patterns are automatically labeled based on the abnormal development process and the effects of subsequent interventions.

[0211] Step 5.4: Implement dynamic updates to the abnormal pattern knowledge base;

[0212] Design an incremental learning algorithm to enable the anomaly classifier to adapt to new anomaly categories;

[0213] Establish an abnormal pattern evaluation mechanism to regularly assess and optimize the knowledge base content;

[0214] Implement knowledge transfer technology to assist in the identification of new anomalies by utilizing the common features of known anomaly patterns.

[0215] The specific implementation of the abnormal pattern self-evolution recognition system includes:

[0216] Anomaly Classifier: Employing an ensemble learning framework, this component combines decision trees, support vector machines, and neural networks. Decision trees handle well-defined anomaly features, support vector machines handle anomaly categories with clear boundaries, and neural networks handle complex nonlinear features. The outputs of each classifier are integrated through a weighted voting mechanism to form the final classification result.

[0217] Online learning module: Employs an incremental learning architecture, which can update model parameters without retraining the entire model.

[0218] For decision trees, the incremental decision tree algorithm is used;

[0219] For support vector machines, use the online support vector machine algorithm;

[0220] For neural networks, gradient accumulation and model fine-tuning methods are used.

[0221] This module includes a memory buffer to store recent outlier samples for model updates. - Pattern Clustering Component: Employs an improved density clustering algorithm capable of handling outlier pattern clustering in high-dimensional feature spaces. This algorithm automatically determines the number of clusters and identifies outliers and emerging anomaly categories based on density information. The clustering component includes a feature extractor, a dimensionality reduction module, and a clustering core. Feature dimensions are reduced using principal component analysis or t-distributed random neighborhood embedding.

[0222] Knowledge Transfer Unit: Based on a meta-learning framework, this unit extracts meta-knowledge from known anomaly categories and applies it to the rapid learning of new anomaly categories. This unit includes a feature extraction layer, a relational network, and a meta-learning optimizer, making it particularly suitable for few-shot learning scenarios.

[0223] In some embodiments, unsupervised or semi-supervised learning methods can be optionally employed to reduce reliance on large amounts of labeled data. For example, contrastive learning or self-supervised learning techniques can be used to pre-train the feature extractor, enhancing the model's ability to generalize to unseen anomalous patterns. Furthermore, multiple anomaly detection algorithms, such as reconstruction-based methods (e.g., autoencoders), density-based methods, and statistical methods, can be integrated to form a more robust anomaly detection system.

[0224] Output: A dynamically updated anomaly pattern recognition system that can identify known anomalies and continuously learn new anomaly patterns.

[0225] like Figure 6 As shown, the variation trends of anomaly identification accuracy between the traditional fixed model and the anomaly pattern self-evolution system proposed in this invention are illustrated during use. It can be seen that the accuracy of the traditional fixed model decreases slightly over time, while the accuracy of the anomaly pattern self-evolution system significantly improves with usage time, increasing from an initial 70% to 91% after 6 months, verifying the "continuous improvement in adaptive capability" mentioned in the results.

[0226] Step 6: Based on the anomaly identification results and the complete causal structure, develop an anomaly diagnosis mechanism to provide an interpretable anomaly root cause diagnosis report through multi-level analysis;

[0227] Specifically, the following steps are included:

[0228] Step 6.1, construct an anomaly tracing and analysis method;

[0229] By backtracking upstream nodes of outlier variables using a cause-effect graph, potential causes can be identified.

[0230] Apply path analysis using causal graphs to calculate the contribution of different causal paths;

[0231] By combining timing information, the order and timeline of occurrence of abnormal events can be determined.

[0232] The specific implementation of the causal path contribution calculation method is as follows:

[0233] Identify all possible paths from potential cause nodes to nodes exhibiting abnormal behavior;

[0234] For each edge on each path, a weight is calculated, with the weight value based on the causal strength and temporal correlation of the edge;

[0235] Calculate the path contribution of each complete path, which is the product or weighted combination of the weights of each edge on the path;

[0236] Sort all paths by contribution to identify the most likely path of anomaly propagation.

[0237] This approach enables the system to distinguish between direct and indirect causes, identify the root cause of an anomaly, and explain its propagation mechanism.

[0238] Step 6.2: Generate multi-level diagnostic interpretations;

[0239] Technical explanation: Detailed description of time-frequency characteristics and causal relationships for professional analysis;

[0240] Management Explanation: A simplified overview of the anomaly and analysis of its main causes, making it easier for decision-makers to understand;

[0241] Operational layer explanation: Specific problem points and suggested operating procedures to guide maintenance personnel in their actions.

[0242] Step 6.3: Construct a visual diagnostic interface;

[0243] Design an interactive cause-effect graph visualization to highlight key anomaly paths;

[0244] Achieve intuitive visualization of time-frequency features and display anomaly characteristics;

[0245] It provides historical comparative analysis, comparing current anomalies with historical cases.

[0246] Step 6.4: Establish an intervention recommendation system;

[0247] Based on the abnormal diagnostic results, combined with counterfactual analysis, targeted intervention suggestions are generated, including measures such as parameter adjustment and equipment inspection.

[0248] It provides a comparative analysis of multiple intervention options and their expected effects.

[0249] In some embodiments, an explanation generation module based on natural language generation technology can be introduced to convert system diagnostic results into natural language descriptions.

[0250] In addition, it can integrate expert knowledge bases, combining causal analysis results and domain expert experience to provide more targeted solution suggestions.

[0251] Output: A detailed anomaly diagnostic report, including anomaly root cause analysis, transmission path, severity assessment, and intervention recommendations.

[0252] like Figure 7 As shown, the energy consumption optimization effect of the smart system proposed in this patent is demonstrated in different types of buildings. With the energy consumption of the traditional system as a baseline of 100%, it can be seen that the smart system achieves a 20% to 30% reduction in energy consumption in different types of buildings such as office buildings, commercial complexes, data centers, and hospitals, while maintaining or improving indoor comfort, thus verifying the "optimized energy utilization" mentioned in the effect of this invention.

[0253] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for dynamic control of building thermal balance based on load forecasting, characterized in that, Includes the following steps: S1: Construct a time-frequency-causal dual-stream analysis network, comprising: a time-frequency analysis branch, consisting of multiple convolutional and pooling layers, employing one-dimensional convolution to extract sequence features along the time axis and two-dimensional convolution to extract local patterns on the time-frequency plane; a causal analysis branch, consisting of a graph convolutional network, with the input being the adjacency matrix and node feature matrix of a preliminary causal graph; and an attention fusion module, consisting of a multi-head self-attention mechanism and a cross-attention mechanism; simultaneously performing time-frequency characteristic analysis and causal relationship analysis on the input heat load time-series data, outputting an enhanced heat load representation after dual-stream analysis; S2: Based on the enhanced heat load representation of S1 output, the fit between the causal structure and the data is evaluated by the Bayesian scoring function. A greedy search strategy is applied to explore the possible causal structure space. Combined with the temporal constraints of the time series, direct causal relationships, indirect causal chains and common cause structures are identified to generate a complete causal graph of the heat load system. S3: Based on the causal graph generated by S2 and the time-frequency features of S1, a causal-enhanced time-frequency representation learning model is constructed. By weighting the time-frequency domain weights with the outputs of multiple causal time-frequency feature extractors, including short-term fluctuation feature extractors, daily cycle feature extractors, long-term trend feature extractors and abnormal pattern feature extractors, the causal-enhanced time-frequency representation is output. S4: Based on the causal enhanced time-frequency representation of S3, construct a structural causal model of the heat load system, remove the incoming edges of the intervened variable through do-calculus operation, calculate the conditional probability distribution of other variables in the system according to the modified causal graph, evaluate the effects of different intervention strategies and select the optimal intervention scheme; S5: Based on the causal enhanced time-frequency representation of S3 and the intervention scheme of S4, an integrated learning framework including decision tree, support vector machine and neural network is constructed. The online update of model parameters is realized through incremental learning algorithm, and new abnormal patterns are identified by density clustering algorithm. The abnormal identification results and updated abnormal pattern library are output. S6: Based on the anomaly identification results of S5 and the cause-effect graph of S2, by identifying all possible paths from potential cause nodes to anomaly manifestation nodes, calculating the weight and path contribution of each path, and generating a multi-level anomaly diagnosis report including technical, management and operational layers.

2. The method according to claim 1, characterized in that, The time-frequency analysis branch in step S1 specifically includes: Select Morlet wavelet or Meyer wavelet as wavelet basis function, perform continuous wavelet transform on indoor temperature, supply and return water temperature, valve opening degree and outdoor meteorological parameters in the input heat load time series data to obtain the time spectrum of each variable; Calculate the coherence coefficient between spectra for different variables, and identify potential associations when the coherence coefficient is greater than a preset threshold; The causal analysis branch determines the conditional independence between variables by calculating the partial correlation coefficient of time series data, applies the PC algorithm to infer the causal direction between variables based on the conditional independence test, and generates the adjacency matrix of the preliminary causal graph. The attention fusion module calculates the correlation weights between time-frequency features and causal features, and achieves mutual enhancement of features through weighted combination.

3. The method according to claim 1, characterized in that, The Bayesian scoring function in step S2 uses either the BIC score or the BDeu score, and the greedy search strategy includes: Initialize an empty cause-effect graph, add or delete edges step by step, and calculate the score value after each modification; The search stops when the score no longer increases, thus obtaining a locally optimal causal structure. The timing constraints include: the timestamp of the cause variable must be earlier than or equal to the timestamp of the result variable; The direct causal relationship is defined as the existence of a direct causal edge connection between variables, the indirect causal chain is defined as a causal path formed through intermediate variables, and the common cause structure is defined as a structure in which multiple variables are affected by the same upstream variable.

4. The method according to claim 1, characterized in that, The causal time-frequency feature extractor in step S3 specifically includes: Short-term fluctuation feature extractor: It adopts a high-pass filter combined with a one-dimensional convolutional neural network, with a convolutional kernel size of 5-15 minutes time span, and includes 3 convolutional layers and 2 pooling layers; Daily Periodic Feature Extractor: Employs a periodic convolution structure, with the convolution kernel reflecting the 24-hour periodicity, and extracts periodic patterns through Fourier transform; Long-term trend feature extractor: Combining wavelet transform and LSTM network, the LSTM has 128 units and contains a 2-layer stacked structure; Anomaly pattern feature extractor: It adopts an autoencoder structure, with the encoder and decoder each containing 3 fully connected layers, and the bottleneck layer has a dimension of 32; The weighted combination is achieved through an adaptive gating mechanism, which dynamically adjusts the weight coefficients of each feature extractor according to the current heat load status.

5. The method according to claim 1, characterized in that, The structural causal model in step S4 is constructed in the following manner: Based on the cause-effect graph, a structural equation is defined between variables, where each endogenous variable is represented as a function of its parent node and the noise term; The parameters of the structural equation are estimated using historical data, employing either the least squares method or maximum likelihood estimation. The do-calculus operation specifically involves: removing all incoming edges of the variable to be intervened from the causal graph, setting the variable as the intervention value, and calculating the posterior distribution of other variables through probabilistic inference; The evaluation system includes energy efficiency indicators, indoor temperature comfort indicators, and system stability indicators, which are weighted and summed to obtain a comprehensive score.

6. The method according to claim 5, characterized in that, The specific implementation of the do-calculus operation includes: Construct a causal graph G before intervention and a causal graph G' after intervention, where G' is obtained by removing all edges from G that point to the intervened variable; Let the variable being intervened be X, and the intervention value be x. For any variable Y, the probability distribution P(Y|do(X=x)) after intervention is calculated as follows: If Y is an ancestor node of X, then P(Y|do(X=x))=P(Y); If Y is a descendant node of X, then the conditional probability is calculated using the d-separation criterion and the chain rule. Monte Carlo simulation was used to generate confidence intervals for estimating the intervention effect using 1000 samples.

7. The method according to claim 1, characterized in that, The ensemble learning framework in step S5 specifically includes: The decision tree classifier uses the C4.5 algorithm to handle anomalous features with well-defined rules; Support vector machines use radial basis kernel functions to handle anomaly categories with clear boundaries; The neural network uses a 3-layer fully connected structure to handle complex nonlinear features; The outputs of each classifier are integrated through a weighted voting mechanism, with the weights determined by the performance of the validation set. The incremental learning algorithm employs the incremental decision tree algorithm for decision trees, the online SVM algorithm for support vector machines, and gradient accumulation and model fine-tuning for neural networks. The density clustering algorithm uses DBSCAN to automatically determine the number of clusters and identify new anomaly categories.

8. The method according to claim 1, characterized in that, The path contribution calculation method in step S6 includes: Identify all paths from potential cause nodes to anomalous behavior nodes using depth-first search or breadth-first search algorithms; Calculate the weight value for each edge, which is equal to the product of the causal strength of that edge and the temporal correlation coefficient; The path contribution of each path is the product of the weights of all edges on the path. The paths are sorted in descending order of their contribution, and the paths with a contribution greater than a threshold are selected as the main anomaly propagation paths. The multi-level diagnostic report is generated through template matching. The technical layer contains detailed parameters and causal paths, the management layer contains an anomaly overview and impact analysis, and the operational layer contains specific maintenance recommendations.

9. The method according to claim 8, characterized in that, The method for calculating the causal strength includes: For linear causal relationships, the causal strength is equal to the path coefficient in the structural equation; For nonlinear causal relationships, the causal strength is determined by calculating the average causal effect or the conditional average causal effect. The time-series correlation coefficient is obtained by calculating the maximum value of the cross-correlation function under different time delays; The path contribution threshold is set to 0.5 times the average contribution of all paths to ensure that only the main anomaly propagation paths are retained.

10. A smart building thermal balance dynamic control system based on load forecasting, used to execute the smart building thermal balance dynamic control method based on load forecasting as described in claim 1, characterized in that, include: The time-frequency-causal dual-stream analysis module includes: a time-frequency analysis submodule, consisting of multiple convolutional and pooling layers, containing one-dimensional and two-dimensional convolutional units; a causal analysis submodule, consisting of a graph convolutional network, containing an adjacency matrix input unit and a node feature processing unit; and an attention fusion submodule, consisting of multi-head self-attention units and cross-attention units. The hierarchical causal discovery module includes: a Bayesian scoring unit for calculating BIC or BDeu scores; a search strategy unit for executing a greedy search algorithm; a temporal constraint unit for ensuring the temporal order of causal relationships; and a causal relationship classification unit for identifying direct causal chains, indirect causal chains, and common cause structures. The causal enhanced time-frequency representation learning module includes: a short-term fluctuation feature extractor, containing a high-pass filter and a 3-layer convolutional network; a daily periodic feature extractor, containing a periodic convolutional layer; a long-term trend feature extractor, containing a wavelet transform unit and a 2-layer LSTM network; an abnormal pattern feature extractor, containing an autoencoder with a 3-layer encoder and a 3-layer decoder; and an adaptive gating unit for dynamically adjusting the weights of each feature extractor. The counterfactual analysis module includes: a structural causal model building unit for defining structural equations between variables; a do-calculus operation unit for removing incoming edges from the causal graph and calculating the probability distribution after intervention; an evaluation index calculation unit for calculating energy efficiency, comfort, and stability indicators; and an intervention strategy selection unit for selecting the optimal solution based on a comprehensive score. The anomaly pattern self-evolution identification module includes: an ensemble classifier containing decision trees, support vector machines, and neural networks; an incremental learning unit for online updating of model parameters; a density clustering unit that uses the DBSCAN algorithm to identify new anomaly categories; and a knowledge base management unit for storing and updating anomaly patterns. The anomaly diagnosis module includes: a path search unit for identifying causal propagation paths; a contribution calculation unit for calculating path weights and path contributions; a multi-level report generation unit for generating diagnostic reports at the technical, management, and operational levels; and a diagnostic visualization unit for displaying causal paths and anomaly characteristics.

Images