Intelligent building heat balance dynamic regulation and control method and system based on load prediction

Abstract

The invention relates to the technical field of intelligent building heat load prediction and heat balance regulation and control, and discloses an intelligent building heat balance dynamic regulation and control method and system based on load prediction, and the method comprises the steps: constructing a time-frequency causal double-flow analysis network, mutual enhancement of the time-frequency characteristics and the causal relationship is realized; designing a hierarchical causal discovery algorithm, and mining a multilevel causal structure; developing a causal enhanced time-frequency representation learning method, and fusing time-frequency and causal information; establishing an intervention decision framework based on anti-fact analysis, and evaluating an intervention effect; an abnormal mode self-evolution recognition system is realized, and new abnormal modes are continuously learned; an interpretable abnormity diagnosis mechanism is developed, and a clear diagnosis report is provided; according to the method, the thermal load prediction accuracy is improved, the abnormal early warning capability is enhanced, the intervention efficiency is improved, the system interpretability is enhanced, and the optimization of the energy utilization efficiency is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent building heat load prediction and heat balance regulation, more specifically, it relates to an intelligent building heat balance dynamic regulation method and system based on load prediction. BACKGROUND

[0002] With the rapid development of intelligent buildings, building heat load prediction and heat balance regulation technology has attracted widespread attention. Current heat load prediction mainly relies on statistical analysis and machine learning methods, which can achieve prediction to some extent, but due to the lack of understanding of the intrinsic causal mechanism of the system, it is difficult to cope with complex and variable actual operating environment.

[0003] Traditional heat balance regulation systems usually use rule-based control strategies, which rely too much on expert experience and are difficult to adapt to dynamic environmental conditions. At the same time, existing anomaly detection and diagnosis methods mainly focus on surface features and cannot effectively trace the root cause of the anomaly, resulting in system response lag and affecting the regulation effect.

[0004] In addition, the existing technology generally lacks explainability, and the decision-making process of the system is a black box to the user, which not only reduces the user's trust, but also increases the difficulty of system maintenance and optimization. The system lacks adaptive learning ability and is difficult to cope with new emerging abnormal patterns, limiting its effectiveness in practical applications. SUMMARY

[0005] The present application provides an intelligent building heat balance dynamic regulation method and system based on load prediction, which solves the technical problems of lack of understanding of deep causal mechanism, insufficient prediction accuracy, lagging anomaly detection, poor explainability and limited adaptive ability in related technologies.

[0006] The present application provides an intelligent building heat balance dynamic regulation method based on load prediction, comprising the following steps:

[0007] S1: Construct a time-frequency-causal dual-flow analysis network, the network includes: a time-frequency analysis branch composed of multiple layers of convolutional layers and pooling layers, using 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 composed of a graph convolution network, the input is the adjacency matrix and node feature matrix of the preliminary causal graph; an attention fusion module composed of multi-head self-attention mechanism and cross-attention mechanism; simultaneously perform time-frequency characteristic analysis and causal relationship analysis on the input heat load time series data, output the enhanced heat load representation after dual-flow analysis;

[0008] S2: Based on the enhanced heat load representation output by S1, evaluate the degree of fitting of the causal structure with the data through the Bayesian scoring function, apply a greedy search strategy to explore the possible causal structure space, identify direct causal relationships, indirect causal chains and common cause structures in combination with the temporal constraints of time series, and 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, build a causal enhanced time-frequency representation learning model, which combines 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, and outputs the causal enhanced time-frequency representation;

[0010] S4: Based on the causal enhanced time-frequency representation of S3, build a structural causal model of the heat load system, remove the incoming edges of the intervened variables 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, build an ensemble learning framework containing decision trees, support vector machines and neural networks, realize online updating of model parameters through incremental learning algorithm, identify new abnormal patterns using density clustering algorithm, and output abnormal identification results and updated abnormal pattern library;

[0012] S6: Based on the abnormal identification results of S5 and the causal graph of S2, calculate the weight of each edge and the total contribution of the path by identifying all possible paths from the latent cause node to the abnormal performance node, and generate a multi-level abnormal diagnosis report containing technical, management and operational layers.

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

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

[0015] Calculate the coherence coefficient between the time-frequency spectra of 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 conditional independence testing, and generates an adjacency matrix of the preliminary causal graph;

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

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

[0019] An empty causal graph is initialized, and edges are gradually added or deleted, and the score value after each modification is calculated;

[0020] The search is stopped when the score value no longer improves, and a local optimal causal structure is obtained;

[0021] The timing constraint includes that 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] Further, the causal time-frequency feature extractor in the step S3 specifically includes:

[0024] Short-term fluctuation feature extractor: adopt high-pass filter combined with one-dimensional convolutional neural network, convolution kernel size is 5-15 minutes time span, including 3 convolution layers and 2 pooling layers;

[0025] Daily cycle feature extractor: adopt periodic convolution structure, convolution kernel reflects 24-hour periodicity, extract periodic pattern through Fourier transform;

[0026] Long-term trend feature extractor: combine wavelet transform and LSTM network, the number of LSTM units is 128, including 2 layers of stacked structure;

[0027] Abnormal pattern feature extractor: adopt autoencoder structure, the encoder and the decoder each include 3 fully connected layers, and the bottleneck layer dimension is 32;

[0028] The weighted combination is realized through an adaptive gating mechanism, and the weight coefficients of each feature extractor are dynamically adjusted according to the current thermal load state.

[0029] Further, the structural causal model in the step S4 is constructed by the following way:

[0030] Define the structural equation between variables based on the causal graph, and each endogenous variable is represented as a function of its parent nodes and noise terms;

[0031] Estimate the parameters of the structural equation through historical data, and adopt least squares method or maximum likelihood estimation;

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

[0033] The evaluation system includes energy consumption efficiency indicators, indoor temperature comfort indicators, and system stability indicators, and a comprehensive score is obtained by weighted summation.

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

[0035] A causal graph G before intervention and a causal graph G' after intervention are constructed, wherein G' is obtained by deleting all edges pointing to the intervened variable from G;

[0036] Let the intervened variable be X, the intervention value be x, and 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 through the d-separation criterion and the chain rule;

[0039] Monte Carlo simulation is used to generate 1000 samples to estimate the confidence interval of the intervention effect.

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

[0041] The decision tree classifier uses the C4.5 algorithm to process abnormal features with clear rules;

[0042] The support vector machine uses a radial basis kernel function to process abnormal classes with clear boundaries;

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

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

[0045] The incremental learning algorithm uses an incremental decision tree algorithm for decision trees, an online SVM algorithm for support vector machines, and a 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 abnormal classes.

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

[0048] All paths from the potential cause node to the abnormal performance node are identified by a depth-first search or breadth-first search algorithm;

[0049] A weight value is calculated for each edge, which is equal to the product of the causal strength and the time correlation coefficient of the edge;

[0050] The total contribution of each path is the product of the weights of the edges on the path;

[0051] The paths are ranked in descending order of contribution, and the paths with a contribution greater than the threshold are selected as the main abnormal propagation paths;

[0052] The multi-level diagnostic report is generated by template matching, the technical layer contains detailed parameters and causal paths, the management layer contains abnormal summary and impact analysis, and the operation layer contains specific maintenance suggestions.

[0053] Further, the calculation method of the causal strength comprises:

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

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

[0056] The time 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 of all path contributions, ensuring that only the main abnormal propagation paths are retained.

[0058] The application also provides a smart building heat balance dynamic regulation system based on load prediction, comprising:

[0059] The time-frequency-causal double-flow analysis module comprises: a time-frequency analysis submodule composed of multiple convolution layers and pooling layers, containing one-dimensional convolution units and two-dimensional convolution units; a causal analysis submodule composed of a graph convolution network, containing an adjacency matrix input unit and a node feature processing unit; an attention fusion submodule composed of a multi-head self-attention unit and a cross-attention unit;

[0060] The hierarchical causal discovery module comprises: a Bayesian scoring unit for calculating BIC score or BDeu score; a search strategy unit for executing a greedy search algorithm; a time sequence constraint unit for ensuring the time sequence logic of the causal relationship; a causal relationship classification unit for identifying direct causal relationship, indirect causal chain and common cause structure;

[0061] The causal enhanced time-frequency representation learning module comprises a short-term fluctuation feature extractor including a high-pass filter and a three-layer convolutional network, a daily cycle feature extractor including a periodic convolutional layer, a long-term trend feature extractor including a wavelet transform unit and a two-layer LSTM network, and an anomaly pattern feature extractor including a self-encoder with a three-layer encoder and a three-layer decoder, and an adaptive gating unit for dynamically adjusting the weights of the feature extractors.

[0062] The counterfactual analysis module comprises a structural causal model construction unit for defining structural equations among variables, a do-calculus operation unit for removing incoming edges in 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 the comprehensive score.

[0063] The anomaly pattern self-evolution identification module comprises an integrated classifier including decision trees, support vector machines and neural networks, an incremental learning unit for online updating model parameters, a density clustering unit using the DBSCAN algorithm to identify new anomaly classes, and a knowledge base management unit for storing and updating anomaly patterns.

[0064] The anomaly diagnosis module comprises a path search unit for identifying causal propagation paths, a contribution calculation unit for calculating path weights and total contribution, a multi-level report generation unit for generating technical, management and operational layer diagnosis reports, and a diagnosis visualization unit for displaying causal paths and anomaly characteristics.

[0065] The beneficial effects of the present application are:

[0066] Through the innovative time-frequency causal dual-flow analysis network, the accuracy of heat load prediction is improved. The present application not only captures the time-frequency features of data, but also deeply understands the internal causal mechanism of the system, making the prediction result more reliable and stable.

[0067] Through the hierarchical causal discovery algorithm and the causal enhanced time-frequency representation learning, a deeper understanding and more accurate control of the heat load system are achieved. The system can automatically discover multi-level causal relationships and effectively integrate these information into the time-frequency representation, improving the prediction and control ability of the system.

[0068] Based on the counterfactual analysis framework and the anomaly pattern self-evolution identification system, the present application improves the system's ability to warn and handle anomalies. The system can discover potential problems in advance and continuously improve its prevention and handling ability by constantly learning and adapting to new anomaly patterns.

[0069] The interpretable anomaly diagnosis mechanism enhances the interpretability and credibility of the system. Through clear causal path analysis and multi-level diagnosis explanation, the system decision-making process is more transparent, which is convenient for managers to understand and operators to execute, and at the same time, the energy utilization efficiency and indoor environmental comfort are improved. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 is a flowchart of a smart building heat balance dynamic regulation method based on load prediction of the present application;

[0071] Figure 2 is a column chart of the accuracy rate of heat load prediction of the present application;

[0072] Figure 3 is a bar chart of the early warning time of the present application;

[0073] Figure 4 is a radar chart of the intervention strategy efficiency of the present application;

[0074] Figure 5 is a column chart of the system interpretability evaluation comparison of the present application;

[0075] Figure 6 is a line chart of the anomaly recognition accuracy rate changing with time of the present application;

[0076] Figure 7 is a column chart of the energy consumption optimization effect of different building types of the present application. DETAILED DESCRIPTION

[0077] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can include changes, modifications, or additions to the functions and arrangements of the elements discussed without departing from the scope of the present disclosure. Various examples can omit, substitute, or add various procedures or components as appropriate or desired. Additionally, the features described herein in connection with one example can also be applicable to other examples.

[0078] In at least one embodiment of the present application, a smart building heat balance dynamic regulation method based on load prediction is disclosed, as shown in Figure 1 includes the following steps:

[0079] Step 1, constructing a time-frequency causal double-flow analysis network, including a time-frequency domain analysis flow and a causal relationship analysis flow, simultaneously processing the time-frequency characteristics and causal relationship of the heat load time series data, outputting preliminary time-frequency characteristics and causal relationship diagram;

[0080] Specifically, the following steps are included:

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

[0082] Preprocess the data, including missing value filling, outlier filtering and time alignment.

[0083] Step 1.2, construct a time-frequency domain analysis flow, use the improved continuous wavelet transform algorithm to process the heat load time series data, and get the time-frequency representation of different time scales and frequency resolution.

[0084] The specific steps are as follows:

[0085] Select a wavelet basis function suitable for the characteristics of heat load data, such as Morlet wavelet or Meyer wavelet;

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

[0087] Calculate the coherence between the time-frequency spectra of different variables to identify potential associations.

[0088] Step 1.3, construct a causal relationship analysis flow, use PC algorithm and conditional independence test to construct a preliminary causal graph structure, representing 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] Apply PC algorithm to infer the causal direction between variables;

[0092] Combine domain knowledge to modify and verify the preliminary causal graph.

[0093] Step 1.4, realize attention fusion mechanism, make the two analysis flows enhance each other;

[0094] The time-frequency analysis flow provides time series characteristics and frequency information for causal discovery, assisting in causal edge identification;

[0095] The causal graph structure provides focus for time-frequency analysis, optimizing time-frequency feature extraction parameters;

[0096] Through cross-attention mechanism, realize information exchange and feature enhancement between the two flows.

[0097] The double-flow analysis network structure includes:

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

[0099] The branch adopts a combination of one-dimensional and two-dimensional convolution, one-dimensional convolution extracts sequence features along the time axis, and two-dimensional convolution extracts local patterns in the time-frequency plane.

[0100] The network depth is 5 layers, each layer is followed by batch normalization and a Rectified Linear Unit (ReLU) activation function after convolution, ensuring the non-linear expression ability of feature extraction.

[0101] Causal analysis branch: composed of a Graph Convolutional Network (GCN), used to handle the causal relationship between variables.

[0102] The GCN contains 3 layers of graph convolution, each layer of graph convolution operation aggregates the information of adjacent nodes to update the node representation. The input is the adjacency matrix and node feature matrix of the preliminary causal graph.

[0103] Attention fusion module: composed of multi-head self-attention mechanism and cross-attention mechanism, realizes the mutual enhancement of the information of the two branches. The attention module calculates the correlation weight between different features, and focuses on the most relevant feature combination.

[0104] In some embodiments, the graph convolutional network can be replaced by a Graph Attention Network (GAT) to enhance the attention ability to important causal relationships.

[0105] In addition, the attention fusion module can adopt different attention mechanisms, for example, optionally adopt a self-attention mechanism based on a Transformer architecture.

[0106] Output: enhanced heat load representation after double-flow analysis, containing time-frequency features and causal structure information, providing a basis for subsequent processing.

[0107] As Figure 2 shown, the heat load prediction accuracy of different prediction methods at different time scales is compared. It can be seen that the causal time-frequency fusion technology proposed in the present application is significantly better than the traditional correlation analysis and single time-frequency analysis method in short-term (1 hour), medium-term (24 hours) and long-term (7 days) prediction, and the prediction accuracy is improved by about 45% on average, verifying the "significant improvement in prediction accuracy" mentioned in the patent technology effect.

[0108] Step 2, based on the time-frequency features and causal relationship graph, a hierarchical causal discovery algorithm is established to discover multi-level causal relationships, including direct causal relationships, indirect causal chains and common cause structures, and output complete causal structures;

[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 goodness-of-fit of causal structures to data - Apply a greedy search strategy to efficiently explore the space of possible causal structures - Incorporate temporal constraints of time-series data to ensure the temporal logic of causal relationships is reasonable

[0111] Step 2.2. Implement multi-level causal relationship mining to distinguish different types of causal relationships.

[0112] Identify direct causal relationships, such as the direct influence of outdoor temperature on water supply temperature.

[0113] Discover indirect causal chains, such as the indirect influence of outdoor temperature on indoor temperature through wall heat storage.

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

[0115] Step 2.3. Combine domain knowledge to revise and verify the causal graph.

[0116] Integrate the basic principles of building thermodynamics as prior knowledge.

[0117] Apply time-lapse-based causal verification methods to test causal hypotheses.

[0118] Use cross-validation techniques to assess the reliability and stability of causal discovery results.

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

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

[0121] Output: A complete causal graph of the heat load system, containing multi-level causal relationships between variables, and annotated with relationship strength and reliability.

[0122] As shown in Figure 3 , the early warning time of different anomaly detection methods for various abnormal situations is compared. It can be seen that the causal enhanced anomaly detection technology proposed in the present invention has a significantly longer early warning time than traditional anomaly detection methods in device failure anomaly, energy consumption anomaly fluctuation, control strategy anomaly and environmental influence anomaly, with an average early warning time increase of about 300%, verifying the "substantial enhancement of anomaly early warning capability" mentioned in the patent technology effect.

[0123] Step 3, using causal structure and time-frequency features, implement causal enhancement time-frequency representation learning, integrate causal information into time-frequency analysis, and output causal enhancement time-frequency representation;

[0124] Specifically, the following steps are included:

[0125] Step 3.1, establish a causal enhancement time-frequency representation learning (CETFR) mathematical model;

[0126] Define the CETFR expression as a comprehensive calculation method that combines the time-frequency domain weight 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 point and frequency point, which reflects the importance of different time-frequency points;

[0129] Calculate the output value of each causal time-frequency feature extractor under the condition of the current data, current time point, current frequency point, and system causal graph;

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

[0131] Sum all the weighted feature extractor output values and multiply them by the time-frequency domain weight value to obtain the final causal enhancement time-frequency representation.

[0132] This representation method can consider both the time-frequency characteristics and causal structure information of the data, and through the weighted combination of multiple specialized feature extractor outputs, it can comprehensively capture the multi-aspect characteristics of the data.

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

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

[0135] Calculate the attention weight based on the relationship strength in the causal graph, focusing on variables and time-frequency regions with strong causal correlation;

[0136] Multiply the preliminary time-frequency representation by the attention weight and process it through a nonlinear activation function to obtain an enhanced feature representation.

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

[0138] Step 3.2, design multiple causal time-frequency feature extractors to address different characteristics of the heat load data;

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

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

[0141] Long-term trend feature extractor, focusing on day to month level heat load changes;

[0142] Abnormal pattern feature extractor, focusing on identifying signal features deviating from normal patterns.

[0143] Step 3.3, optimize time-frequency analysis parameters based on causal structure;

[0144] Use important nodes and edges in causal graph to guide the focus area of time-frequency analysis;

[0145] Adjust the time window and frequency resolution of wavelet transform dynamically according to causal relationship strength;

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

[0147] Step 3.4, realize adaptive fusion of causal time-frequency features;

[0148] Design adaptive weight calculation mechanism, dynamically adjust the weight of each feature extractor according to the current heat load state;

[0149] Use attention mechanism to weight the importance of different time-frequency points and features;

[0150] Construct multi-level fusion strategy, gradually integrate from low-level time-frequency features to high-level semantic features.

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

[0152] Short-term fluctuation feature extractor: adopt high-pass filter combined with one-dimensional convolutional neural network structure, convolution kernel size set to 5 to 15 minutes time span, containing 3 convolution layers and 2 pooling layers, specially capture fast changing heat load fluctuation features.

[0153] Input is normalized time series data, output is feature vector representing short-term fluctuation characteristics.

[0154] Daily cycle feature extractor: adopt periodic convolution structure, convolution kernel design reflects 24-hour periodicity, extract periodic patterns through Fourier transform and harmonic analysis.

[0155] This extractor contains periodic padding layer, periodic convolution layer and global pooling layer, specially focusing on daily load change pattern.

[0156] Long-term trend feature extractor: Combining wavelet transform and Long Short-Term Memory (LSTM), the long-term trend features are extracted from down-sampled long time series data.

[0157] The LSTM unit number is 128, containing 2 layers of stacked structure, and the slow changing trend at the scale of day to month is captured by setting the time step.

[0158] Abnormal pattern feature extractor: Using autoencoder structure, containing encoder and decoder two parts. The encoder compresses the input data to low-dimensional latent space, and the decoder tries to reconstruct the original data.

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

[0160] The adaptive fusion network adopts a gating mechanism, including:

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

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

[0163] Context-aware gating mechanism, dynamically adjusting the weight according to the current thermal load state and historical data;

[0164] Residual connection structure, ensuring that important information is not lost in the fusion process.

[0165] In some embodiments, different sequence models other than LSTM can be used, such as Gated Recurrent Unit (GRU) or bidirectional LSTM, to enhance the long-term trend feature extraction capability.

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

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

[0168] For example, Figure 4As shown, the traditional intervention strategy and the intervention strategy based on counterfactual analysis proposed in this patent are compared in terms of response time, energy consumption, stability, energy saving effect and user comfort. It can be seen that the counterfactual analysis intervention is significantly better than the traditional intervention strategy in all dimensions, and the overall efficiency is improved by about 60%, which verifies the "intervention efficiency is significantly improved" mentioned in the patent technology effect.

[0169] Step 4, based on the time-frequency representation of causal enhancement, construct the counterfactual analysis framework, evaluate the intervention strategy effect through do-calculus operation, output the optimal intervention scheme;

[0170] Specifically, the following steps are included:

[0171] Step 4.1, construct the structural causal model (SCM) of the thermal load system;

[0172] Based on the discovered causal graph, define the functional relationship between variables;

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

[0174] Ensure that the SCM can accurately simulate system behavior through model validation.

[0175] Step 4.2, realize random intervention simulation and evaluate intervention effect;

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

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

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

[0179] Do-calculus operation is an intervention algorithm, which is implemented as follows:

[0180] Remove all incoming edges of the intervened variable in the causal graph (i.e. the causal influence of other variables on this variable), indicating that the variable is no longer affected by its original factors, but is artificially set to a specific value;

[0181] Calculate the conditional probability distribution of other variables in the system according to the modified causal graph;

[0182] Generate the overall state prediction of the system after intervention through probability reasoning or Monte Carlo simulation method.

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

[0184] Step 4.3, develop intervention effect evaluation index system;

[0185] Define energy efficiency indicators to assess the impact of intervention on energy consumption;

[0186] Design comfort indicators to assess the impact of intervention on indoor environment;

[0187] Build stability indicators to assess the impact of intervention on system balance;

[0188] Calculate the comprehensive score to balance the multi-dimensional goals.

[0189] Step 4.4, establish optimal intervention strategy selection algorithm;

[0190] Based on the evaluation index, sort different intervention strategies;

[0191] Consider the implementation cost and technical feasibility, filter out unsuitable intervention options;

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

[0193] In some embodiments, Markov decision process or partially observable Markov decision process can be used to model the intervention decision problem to handle the case of partially observable system state.

[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: Optimal intervention strategy for current thermal load state, including intervention point, intervention parameter and expected effect.

[0196] As shown in Figure 5 , the interpretability of the traditional system and the causal path analysis system proposed by this method is compared 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 is significantly better than the traditional system in all dimensions, especially in the decision-making process transparency, which verifies the "overall enhancement of system interpretability" mentioned in the patent technology effect.

[0197] Step 5, according to the time-frequency representation of causal enhancement and the intervention scheme, realize the abnormal pattern self-evolution recognition system, through the cooperation of multiple modules to continuously learn new abnormal patterns, output abnormal recognition results;

[0198] Specifically, the following steps are included:

[0199] Step 5.1, establish a basic abnormal pattern library;

[0200] Collect and label common thermal load abnormal patterns such as equipment failure, abnormal energy consumption, etc.

[0201] Extract time-frequency features and causal features of each abnormal pattern to form feature templates.

[0202] Build an initial abnormality classifier to identify known types of abnormalities based on template matching.

[0203] Step 5.2, implement online anomaly detection algorithm;

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

[0205] Calculate the deviation of the current representation from the normal pattern to quantify the degree of abnormality;

[0206] Apply a multi-level threshold strategy to implement different levels of abnormality warning.

[0207] Step 5.3, build a new abnormal pattern learning mechanism;

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

[0209] When similar unknown abnormalities repeatedly occur, cluster and summarize the patterns;

[0210] According to the development process of the anomaly and the subsequent intervention effect, automatically label new abnormal patterns.

[0211] Step 5.4, implement dynamic updating of abnormal pattern knowledge base;

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

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

[0214] Implement knowledge transfer technology to use the common features of known abnormal patterns to assist in identifying new abnormalities.

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

[0216] Abnormality classifier: Use an ensemble learning framework combining decision trees, support vector machines, and neural networks. Decision trees are used to handle well-defined abnormal features, support vector machines are used to handle clear boundary abnormal categories, and neural networks are used to 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: Use an incremental learning architecture that can update model parameters without retraining the entire model.

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

[0219] For support vector machines, an online support vector machine algorithm is used;

[0220] For neural networks, a gradient accumulation and model fine-tuning method is used.

[0221] The module contains a memory buffer that stores recent anomaly samples for model updates. - Pattern clustering component: uses an improved density clustering algorithm that can handle anomaly pattern clustering in high-dimensional feature space. The algorithm automatically determines the number of clusters and identifies outliers and emerging anomaly classes based on density information. The clustering component includes a feature extractor, a dimensionality reduction module, and a clustering core, and the feature dimension is reduced by principal component analysis or t-distributed stochastic neighbor embedding.

[0222] Knowledge transfer unit: based on the meta-learning framework, extracts meta-knowledge from known anomaly classes and applies it to the rapid learning of new anomaly classes. The unit includes a feature extraction layer, a relationship network, and a meta-learning optimizer, and is particularly suitable for few-shot learning scenarios.

[0223] In some embodiments, unsupervised or semi-supervised learning methods are optionally used to reduce dependence on large amounts of labeled data. For example, contrastive learning or self-supervised learning techniques can be used to pretrain feature extractors, enhancing the model's ability to generalize to unseen anomaly patterns. In addition, multiple anomaly detection algorithms can be integrated, such as reconstruction-based methods (e.g., autoencoders), density-based methods, and statistical-based methods, 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 emerging anomaly patterns.

[0225] As shown in Figure 6 The change trend of anomaly recognition accuracy in the use process of the traditional fixed model and the anomaly pattern self-evolution system proposed by the present application is shown. 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 time, from the initial 70% to 91% after 6 months, verifying the "continuous improvement of adaptive ability" mentioned in the effect.

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

[0227] Specifically, the following steps are included:

[0228] Step 6.1, construct an anomaly root cause analysis method;

[0229] Based on the upstream nodes of the anomaly variable in the causal graph, find the potential causes;

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

[0231] Determine the sequence and timeline of abnormal events by combining temporal information.

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

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

[0234] Calculate the weight of each edge on each path, and the weight value is based on the causal strength and temporal relevance of the edge;

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

[0236] Sort all paths by contribution and identify the most likely abnormal propagation path.

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

[0238] Step 6.2, implement multi-level diagnostic explanation generation;

[0239] Technical layer explanation: detailed time-frequency features and causal relationship description for professional analysis;

[0240] Management layer explanation: simplified abnormal overview and main cause analysis for decision makers to understand;

[0241] Operational layer explanation: specific problem points and recommended operation steps to guide maintenance personnel.

[0242] Step 6.3, build a visual diagnostic interface;

[0243] Design interactive causal graph visualization to highlight key abnormal paths;

[0244] Implement intuitive visual representation of time-frequency features to display abnormal characteristics;

[0245] Provide historical comparison analysis to compare the current anomaly with historical cases.

[0246] Step 6.4, establish an intervention suggestion system;

[0247] Generate targeted intervention suggestions based on anomaly diagnosis results, including parameter adjustment and equipment inspection measures, etc.

[0248] Provide multiple intervention options and comparative analysis of their expected effects.

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

[0250] In addition, an expert knowledge base can be integrated to provide more targeted solution suggestions in combination with causal analysis results and domain expert experience.

[0251] Output: Detailed abnormality diagnosis report, including abnormality root analysis, propagation path, severity assessment, and intervention suggestions.

[0252] As Figure 7 shown, the energy consumption optimization effect of the intelligent system proposed in this patent in different types of buildings is demonstrated. Taking the energy consumption of the traditional system as 100% benchmark, it can be seen that the intelligent system has achieved 20% to 30% energy consumption reduction in office buildings, commercial complexes, data centers, and hospitals, etc. different types of buildings, while maintaining or improving indoor comfort, verifying the "optimizing energy utilization" mentioned in the effect of the invention.

[0253] The embodiments of the present application are described above, but the embodiments are not limited to the specific implementation described above, which is only illustrative and not limiting. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A method for dynamic control of heat balance in smart buildings based on load forecasting, characterized in that: The following steps are involved: S1: Construct a time-frequency-causal dual-stream analysis network, which includes: a time-frequency analysis branch consisting of multiple convolutional layers and pooling layers, using 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, taking as input the adjacency matrix and node feature matrix of the preliminary causal graph; an attention fusion module consisting of a multi-head self-attention mechanism and a cross-attention mechanism; the network simultaneously performs time-frequency characteristic analysis and causal relationship analysis on the input heat load time series data, and outputs an enhanced heat load representation after dual-stream analysis; S2: Based on the enhanced heat load representation output by S1, the Bayesian scoring function is used to evaluate the fit between the causal structure and the data. A greedy search strategy is applied to explore the space of possible causal structures. 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 diagram for 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. The model outputs a causal-enhanced time-frequency representation by weightedly combining the time-frequency domain weights with the outputs of multiple causal time-frequency feature extractors, including a short-term fluctuation feature extractor, a daily cycle feature extractor, a long-term trend feature extractor, and an abnormal pattern feature extractor. S4: Based on the causal enhanced time-frequency representation of S3, a structural causal model of the heat load system is constructed. The input edges of the intervened variables are removed through do-calculus operations. The conditional probability distribution of other variables in the system is calculated according to the modified causal graph. The effects of different intervention strategies are evaluated and the optimal intervention plan is selected. S5: Based on the causal enhanced time-frequency representation of S3 and the intervention plan of S4, an integrated learning framework consisting of decision trees, support vector machines, and neural networks is constructed. The model parameters are updated online through an incremental learning algorithm. A density clustering algorithm is used to identify new abnormal patterns. The anomaly identification results and the updated abnormal pattern library are output. S6: Based on the anomaly identification results of S5 and the causal diagram of S2, by identifying all possible paths from the potential cause node to the anomaly manifestation node, calculating the weight of each path and the total contribution of the path, a multi-level anomaly diagnosis report covering the technical layer, management layer, and operation layer is generated.

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 the wavelet basis function, perform continuous wavelet transform on the indoor temperature, supply and return water temperature, valve opening, and outdoor meteorological parameters in the input heat load time series data, and obtain the time spectrum of each variable; Calculate the coherence coefficient between the time spectra of different variables, and identify potential correlation 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 an adjacency matrix of a 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 adopts BIC scoring or BDeu scoring, and the greedy search strategy includes: Initialize an empty causal graph, gradually add or delete edges, and calculate the score value after each modification; When the score value no longer increases, the search stops and the local optimal causal structure is obtained; 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 the 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, wherein The causal time-frequency feature extractor in step S3 specifically includes: Short-term fluctuation feature extractor: uses a high-pass filter combined with a one-dimensional convolutional neural network. The convolution kernel size is 5-15 minutes long, and it contains 3 convolutional layers and 2 pooling layers. Daily cycle feature extractor: uses a periodic convolution structure with a 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 number of LSTM units is 128, including a 2-layer stacked structure; Abnormal pattern feature extractor: uses an autoencoder structure. The encoder and decoder each contain 3 fully connected layers, and the bottleneck layer dimension is 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 state.

5. The method according to claim 1, wherein The structural causal model in step S4 is constructed in the following way: The structural equations between variables are defined based on the causal graph, where each endogenous variable is represented as a function of its parent node and the noise term; Estimate the parameters of the structural equation using historical data, using the least squares method or maximum likelihood estimation; The do-calculus operation is specifically as follows: all incoming edges of the intervened variable are removed from the causal graph, the variable is set to the intervention value, and the posterior distribution of other variables is calculated through probabilistic reasoning; The evaluation system includes energy efficiency index, indoor temperature comfort index and system stability index, and a comprehensive score is obtained through weighted summation.

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 the intervention and a causal graph G' after the intervention, where G' is obtained by deleting all edges pointing to the intervened variable from G; Let the intervened variable 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 of X, then P(Y|do(X=x))=P(Y); If Y is a descendant node of X, the conditional probability is calculated using the d-separation criterion and the chain rule; Monte Carlo simulations were used to generate confidence intervals for 1000 sample estimates of the intervention effect.

7. The method according to claim 1, characterized in that The integrated learning framework in step S5 specifically includes: The decision tree classifier uses the C4.5 algorithm to process abnormal features with clear rules; The support vector machine uses the radial basis kernel function to handle abnormal categories with clear boundaries; The neural network adopts a 3-layer fully connected structure to handle complex nonlinear features; The outputs of each classifier are integrated through a weighted voting mechanism, and the weights are determined by the performance of the validation set; The incremental learning algorithm adopts an incremental decision tree algorithm for decision trees, an 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 abnormal manifestation nodes through 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 the edge and the temporal correlation coefficient; The total contribution of each path is the product of the weights of each edge on the path; Arrange the paths in descending order of contribution, and select the path with a contribution greater than the threshold as the main anomaly propagation path; Multi-level diagnostic reports are generated through template matching, with the technical level containing detailed parameters and causal paths, the management level containing anomaly overview and impact analysis, and the operational level containing specific maintenance recommendations.

9. The method according to claim 8, characterized in that The calculation method of the causal strength includes: For linear causality, causal strength is equal to the path coefficient in the structural equation; For nonlinear causal relationships, causal strength is determined by calculating the average causal effect or conditional average causal effect; The temporal 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 mean of all path contributions to ensure that only the main abnormal propagation paths are retained.

10. A smart building heat balance dynamic control system based on load forecasting, characterized in that: include: The time-frequency-causal dual-stream analysis module includes: a time-frequency analysis submodule, which consists of multiple convolutional layers and pooling layers, including one-dimensional convolution units and two-dimensional convolution units; a causal analysis submodule, which consists of a graph convolutional network, including an adjacency matrix input unit and a node feature processing unit; and an attention fusion submodule, which consists of a multi-head self-attention unit and a cross-attention unit. The hierarchical causal discovery module includes: a Bayesian scoring unit for calculating BIC scores 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 relationships, indirect causal chains, and common cause structures. The causal enhanced time-frequency representation learning module includes: a short-term fluctuation feature extractor with a high-pass filter and a three-layer convolutional network; a daily cycle feature extractor with a periodic convolutional layer; a long-term trend feature extractor with a wavelet transform unit and a two-layer LSTM network; an abnormal pattern feature extractor with an autoencoder with a three-layer encoder and a three-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 construction unit for defining the structural equations between variables; a do-calculus operation unit for removing incoming edges from the causal graph and calculating the probability distribution after the 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 the comprehensive score. The abnormal pattern self-evolution recognition module includes: an integrated classifier, including 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 abnormal categories; and a knowledge base management unit for storing and updating abnormal patterns. The anomaly diagnosis module includes: a path search unit for identifying causal propagation paths; a contribution calculation unit for calculating path weights and total 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