Campus air conditioner energy consumption prediction method based on AIGC reasoning

Abstract

The invention discloses a campus air conditioner energy consumption prediction method based on AIGC reasoning, and the method comprises the following steps: collecting and preprocessing multi-source environment data, and forming a multi-source energy consumption data set; key variables are extracted, and a semantic state vector set is generated; generating a virtual energy consumption semantic sample, and forming a fused energy consumption semantic sample set; constructing an energy consumption dynamic equation, and outputting an energy consumption state response sequence; establishing an energy consumption time sequence prediction function, predicting an energy consumption change trend, and generating a preliminary energy consumption prediction result; comparing with real-time monitoring data, calculating a prediction error value, executing a joint self-adaptive optimization action, and outputting a model parameter set after joint self-adaptive optimization; performing semantic re-reasoning operation to generate a semantic enhanced energy consumption prediction result; and performing visual output, and generating an energy consumption trend curve and a load change report. According to the method, AIGC and the liquid neural network are fused, intelligent campus energy consumption prediction is realized, and the method has the advantages of high precision and self-adaption.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-source energy consumption data fusion analysis, and particularly relates to a campus air conditioner energy consumption prediction method based on AIGC reasoning. BACKGROUND

[0002] The existing campus building energy consumption prediction methods are mostly based on traditional time series models or deep learning structures, such as long short-term memory networks, gated recurrent units and Transformer models based on attention mechanisms. These methods usually rely on historical energy consumption data for one-way fitting, lack modeling capability for semantic level energy consumption factors, and are difficult to express the complex correlation between environmental changes, personnel activities and equipment operating states. When the energy consumption data is sparse, missing or unevenly distributed, the prediction accuracy is significantly reduced, and the model cannot dynamically adjust according to real-time changes, resulting in prediction lag and energy consumption management response delay.

[0003] In view of the above problems, the existing technology has not formed an energy consumption prediction method that can simultaneously combine multi-source Internet of Things data, semantic generation mechanism and continuous time neural dynamic modeling. Traditional models only perform correlation learning at the data statistical level, cannot generate virtual energy consumption samples in the semantic space for feature expansion, and cannot realize adaptive updating of model parameters with scene changes. The present application introduces an AIGC semantic generation mechanism and improves the liquid neural network to build a semantic-driven dynamic reasoning and joint optimization closed loop, thereby fundamentally solving the problems of insufficient prediction accuracy and timeliness.

[0004] Therefore, how to provide a campus air conditioner energy consumption prediction method based on AIGC reasoning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a campus air conditioner energy consumption prediction method based on AIGC reasoning. The present application combines AIGC and liquid neural networks to realize intelligent prediction of campus energy consumption, and has the advantages of high precision and adaptability.

[0006] According to the campus air conditioner energy consumption prediction method based on AIGC reasoning, the following steps are included:

[0007] Collecting multi-source environmental data of the air conditioning system in the campus building group, performing time synchronization, anomaly elimination, interpolation completion and normalization processing on the data, and forming a structured multi-source energy consumption dataset;

[0008] Extracting key variables and generating a semantic state vector set using a feature encoding network;

[0009] Inputting the semantic state vector set into an energy consumption semantic generation mechanism to generate virtual energy consumption semantic samples consistent with the real distribution, and forming a fusion energy consumption semantic sample set.

[0010] inputting the fusion energy consumption semantic sample set into an improved liquid neural network, constructing a continuous-time energy consumption dynamic equation, and outputting an energy consumption state response sequence;

[0011] based on the energy consumption state response sequence, establishing an energy consumption time sequence prediction function, predicting the energy consumption change trend in a future period, and generating a preliminary energy consumption prediction result;

[0012] comparing the preliminary energy consumption prediction result with real-time monitoring data in the multi-source energy consumption data set, calculating a prediction error value, performing a joint adaptive optimization action, and outputting a model parameter set after joint adaptive optimization;

[0013] inputting the optimized energy consumption state response sequence into an energy consumption semantic generation mechanism, performing semantic re-reasoning operation under the constraint of the corrected semantic generation parameters, and generating a semantic enhanced energy consumption prediction result;

[0014] visualizing the final energy consumption prediction result and generating an energy consumption trend curve and a load change report.

[0015] Optionally, the generation of the semantic state vector set comprises:

[0016] time synchronization, anomaly elimination and missing value completion are performed on the original sensor data in the multi-source energy consumption data set, data of different sources are aligned under a unified time scale, data gaps are repaired through linear interpolation and median smoothing method, and a time-continuous multi-source feature sequence is obtained;

[0017] the multi-source feature sequence is input into a feature standardization module, and normalization processing is performed on numerical features of different dimensions to construct a multi-dimensional input feature matrix;

[0018] the feature matrix is input into a feature encoding network with multi-scale convolution and recursive structure, short-term energy consumption change features are extracted through local convolution units, long-term change trends are captured through time recursion units, and an intermediate feature representation sequence describing the dynamic change pattern of energy consumption is generated;

[0019] semantic mapping operation is performed on the basis of the intermediate feature representation sequence, feature importance weights are calculated according to the contribution of each input feature to energy consumption change, and weighted fusion is performed according to the semantic relevance between features and energy consumption, and the semantic embedding representation is obtained by weighted summation of all features at each time step.

[0020] the semantic embedding representation is input into a time fusion unit to perform time continuity correction, smooth interpolation and dynamic adjustment are performed according to the semantic change amplitude of adjacent time points, and a time-continuous semantic state vector set is output.

[0021] Optionally, the formation of the fusion energy consumption semantic sample set comprises:

[0022] inputting the semantic state vector set into an energy consumption semantic generation mechanism to generate an initial virtual energy consumption semantic sample set;

[0023] performing distribution consistency correction on the initial virtual energy consumption semantic sample set, calculating a difference degree of the virtual energy consumption semantic sample set and the multi-source energy consumption data set in statistical features, and outputting a corrected virtual energy consumption semantic sample set when the difference degree converges to a set threshold;

[0024] performing time continuity correction on the corrected virtual energy consumption semantic sample set to output a time-continuous virtual energy consumption semantic sample set;

[0025] performing semantic-level fusion of the time-continuous virtual energy consumption semantic sample set and the multi-source energy consumption data set, performing feature mapping and weighted superposition in the same time step or adjacent time steps, dynamically allocating fusion weights according to sample confidence, and obtaining a fused energy consumption semantic sample set.

[0026] Optionally, the generation of the energy consumption state response sequence comprises:

[0027] inputting the fused energy consumption semantic sample set into an improved liquid neural network, initializing an input layer of the liquid neural network according to feature vectors of each time step in the fused energy consumption semantic sample set, and establishing an energy consumption dynamic equation of continuous time;

[0028] introducing a semantic driving regulation term into the energy consumption dynamic equation of the liquid neural network to output a neural state sequence after semantic driving regulation;

[0029] performing an adaptive time constant updating mechanism on the neural state sequence to output a neural state sequence after time constant correction;

[0030] inputting the neural state sequence into a network output layer to obtain an energy consumption state response sequence after normalization processing.

[0031] Optionally, the generation of the preliminary energy consumption prediction result comprises:

[0032] establishing an energy consumption time sequence prediction function based on the energy consumption state response sequence output by the improved liquid neural network, and forming a time-dependent mapping relationship between the input and the prediction output by setting a time step extension window;

[0033] calculating the dependence relationship between each time step in the energy consumption state response sequence and extracting an energy consumption time sequence feature matrix, and outputting the energy consumption time sequence feature matrix;

[0034] inputting the energy consumption time sequence feature matrix into the energy consumption time sequence prediction function to perform prediction reasoning, and outputting an energy consumption change trend sequence;

[0035] According to the energy consumption change trend sequence, the energy consumption change amplitude and direction of the future period are calculated, and a preliminary energy consumption prediction result is generated.

[0036] Optionally, the generation of the jointly self-adaptive optimized model parameter set comprises:

[0037] The preliminary energy consumption prediction result is time-synchronized with the real-time monitoring data in the multi-source energy consumption data set, and the prediction value and the measured energy consumption value at the same time step are differentially calculated to obtain a prediction error value.

[0038] The prediction error change rate is calculated, and the current energy consumption scene change trend is judged according to the error change rate. When the error change rate is greater than a set threshold, it is determined as an energy consumption mutation scene, and when the error change rate is lower than the threshold, it is determined as an energy consumption smooth scene.

[0039] According to the prediction error change rate, the time constant of the improved liquid state neural network is adjusted and improved, so that the time constant is inversely proportional to the error change rate. When the energy consumption mutation scene appears, the time constant is shortened to improve the response speed. When the energy consumption smooth scene appears, the time constant is lengthened to maintain the stability of the network output. The neural state parameters after time constant adjustment are output.

[0040] Based on the prediction error value and the prediction error change rate, a joint self-adaptive optimization mechanism is established. According to the minimum sum of squared errors objective function, the weight matrix parameters of the improved liquid state neural network are updated by gradient, and the prediction error information is input into the energy consumption semantic generation mechanism. According to the prediction error distribution characteristics, the semantic generation parameters are corrected.

[0041] The time constant, the weight matrix parameters of the improved liquid state neural network after joint self-adaptive optimization, and the corrected semantic generation parameters of the energy consumption semantic generation mechanism are integrated to form a jointly self-adaptive optimized model parameter set.

[0042] Optionally, the generation of the semantic enhancement type energy consumption prediction result comprises:

[0043] The jointly self-adaptive optimized model parameter set is called, and the model parameter set is loaded into the improved liquid state neural network and the energy consumption semantic generation mechanism.

[0044] The energy consumption state response sequence output by the improved liquid state neural network is input into the energy consumption semantic generation mechanism, and the semantic reasoning operation is performed under the constraint of the corrected semantic generation parameters. The semantic enhancement type energy consumption prediction result is generated based on the semantic condition.

[0045] The semantic enhancement type energy consumption prediction result is compared with real-time monitoring data in the multi-source energy consumption data set, a new prediction error value is calculated according to the difference between the prediction result and the real-time monitoring data, and the time constant and weight matrix parameters of the improved liquid neural network and the semantic generation parameters of the energy consumption semantic generation mechanism are iteratively updated online according to the prediction error value, to form a dynamically updated model parameter set.

[0046] Optionally, the generation of the energy consumption trend curve and the load change report comprises:

[0047] The semantic enhancement type energy consumption prediction result is compared with real-time monitoring data in the multi-source energy consumption data set, a new prediction error value is calculated according to the difference between the prediction result and the real-time monitoring data, and the time constant and weight matrix parameters of the improved liquid neural network and the semantic generation parameters of the energy consumption semantic generation mechanism are iteratively updated online according to the prediction error value, to form a dynamically updated model parameter set.

[0048] The final energy consumption prediction result is data-arranged and time-series expanded, and an energy consumption prediction sequence of a continuous time period is generated according to the energy consumption values of each time step in the prediction result;

[0049] According to the energy consumption prediction sequence, the energy consumption change rate per unit time and the load fluctuation amplitude are calculated, and an energy consumption trend curve is drawn.

[0050] The energy consumption trend curve is statistically analyzed, and the energy consumption peak period, the load fluctuation frequency and the change stability index are extracted, and a load change report is generated accordingly;

[0051] The energy consumption trend curve and the load change report are visually output, and the energy consumption prediction curve, the load fluctuation interval and the report abstract are displayed on the time axis to form a visual display output of the final energy consumption prediction result.

[0052] The present application has the following advantages:

[0053] The present application introduces an AIGC semantic generation mechanism and an improved liquid neural network in the energy consumption prediction process, realizes intelligent, semantic and dynamic modeling of the air conditioning energy consumption of the campus building group, unifies the collection and preprocessing of multi-source energy consumption data, constructs a structured data set containing environmental factors, personnel density and equipment state and other multi-dimensional characteristics, so that the energy consumption prediction is no longer limited to the fitting of single historical data, but can integrate the global energy consumption characteristics in the Internet of Things environment. On this basis, a feature coding network with multi-scale convolution and recursive structure is used to extract key feature variables and generate semantic state vectors, so as to realize semantic modeling of energy consumption influencing factors and make the model have the expression ability to complex scene factors.

[0054] The virtual energy consumption semantic sample is generated under the semantic condition constraint through the AIGC semantic generation mechanism, and is fused with the real observation data at the semantic level, which overcomes the limitations of sparse, missing and noise interference of energy consumption data. The mechanism realizes the expansion of the energy consumption sample in the semantic space, so that the model can maintain high prediction stability and generalization ability when the sample is insufficient or the data distribution is uneven. The improved liquid neural network introduces a semantic driving control term and an adaptive time constant updating mechanism in the traditional liquid unit, so that the energy consumption dynamic equation can adaptively adjust the neural state evolution rate according to the change of semantic characteristics, thereby realizing the continuous and smooth updating of energy consumption response with the change of scene semantics. This mechanism significantly improves the response ability of the model to sudden energy consumption changes and environmental disturbances.

[0055] The present application realizes the cooperative optimization of the semantic generation mechanism and the liquid neural network by establishing an energy consumption time sequence prediction function and a joint adaptive optimization mechanism. When there is an error between the prediction result and the real-time monitoring data, the system can correct the network weight parameters and the semantic generation parameters based on the error feedback, forming a closed-loop self-learning structure, so that the model can continuously optimize and dynamically stabilize itself in the long-term operation. Finally, the semantic enhanced energy consumption prediction result is generated through semantic re-reasoning, and the energy consumption trend curve and load change report are output in the visualization stage, realizing the observability, explainability and verifiability of energy consumption changes. The present application improves the accuracy, stability and real-time performance of the campus air conditioning system energy consumption prediction as a whole, and provides reliable data support and intelligent decision basis for energy saving regulation and energy consumption optimization of the smart campus. BRIEF DESCRIPTION OF DRAWINGS

[0056] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, and do not constitute a limitation of the present application. In the drawings:

[0057] Figure 1 A flowchart of a campus air conditioning energy consumption prediction method based on AIGC reasoning proposed by the present application;

[0058] Figure 2 An energy consumption time sequence prediction function establishment and prediction reasoning flowchart in a campus air conditioning energy consumption prediction method based on AIGC reasoning proposed by the present application. DETAILED DESCRIPTION

[0059] The present application will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0060] REFERENCE Figure 1 and Figure 2 A campus air conditioning energy consumption prediction method based on AIGC reasoning, comprising the following steps:

[0061] Collecting multi-source environment data of air conditioning systems in a campus building group, time synchronizing, excluding outliers, interpolating and completing data, and normalizing the data to form a structured multi-source energy consumption dataset;

[0062] Extracting key variables and generating a semantic state vector set using a feature encoding network;

[0063] Inputting the semantic state vector set into an energy consumption semantic generation mechanism to generate virtual energy consumption semantic samples consistent with the real distribution, forming a fusion energy consumption semantic sample set;

[0064] Inputting the fusion energy consumption semantic sample set into an improved liquid neural network to construct a continuous-time energy consumption dynamic equation and output an energy consumption state response sequence;

[0065] Based on the energy consumption state response sequence, an energy consumption time series prediction function is established to predict the energy consumption trend in the future period and generate a preliminary energy consumption prediction result;

[0066] Comparing the preliminary energy consumption prediction result with the real-time monitoring data in the multi-source energy consumption dataset, calculating the prediction error value, performing joint adaptive optimization action, and outputting the model parameter set after joint adaptive optimization;

[0067] Inputting the optimized energy consumption state response sequence into the energy consumption semantic generation mechanism to perform semantic re-reasoning operation under the constraint of the corrected semantic generation parameters, and generating a semantic enhanced energy consumption prediction result;

[0068] Visualizing the final energy consumption prediction result to generate an energy consumption trend curve and a load change report.

[0069] In the embodiment, the generation of the semantic state vector set includes:

[0070] Time synchronizing, excluding outliers, and completing missing values of the original sensor data in the multi-source energy consumption dataset, aligning data from different sources under a unified time scale, repairing data gaps through linear interpolation and median smoothing method, and obtaining a time-continuous multi-source feature sequence;

[0071] Inputting the multi-source feature sequence into a feature normalization module to perform normalization processing on numerical features of different dimensions, and constructing a multi-dimensional input feature matrix, wherein each row corresponds to a feature vector at a time step;

[0072] The feature matrix is input into a feature encoding network of a multi-scale convolution and a recursive structure, the feature encoding network comprising a multi-scale convolution feature extraction layer, a recursive time feature aggregation layer, a feature attention encoding layer, and a semantic state vector generation layer, short-term energy consumption change features are extracted through a local convolution unit, long-term change trends are captured through a time recursion unit, and an intermediate feature representation sequence describing energy consumption dynamic change patterns is generated;

[0073] On the basis of the intermediate feature representation sequence, a semantic mapping operation is performed, feature importance weights are calculated according to the contribution of each input feature to energy consumption change, and weighted fusion is performed according to the semantic correlation between the features and the energy consumption, the semantic embedding representation is obtained by weighted summation of all features at each time step, and each component of the semantic embedding representation corresponds to an environmental semantic component, a building use semantic component, and a personnel activity semantic component respectively;

[0074] The semantic embedding representation is input into a time fusion unit to perform time continuity correction, smooth interpolation and dynamic adjustment are performed according to the semantic change amplitude of adjacent time points, and a set of time-continuous semantic state vectors is output, the set of time-continuous semantic state vectors is used to represent the dynamic semantic conditions of energy consumption influencing factors.

[0075] In the embodiment, the formation of the fused energy consumption semantic sample set comprises:

[0076] The set of time-continuous semantic state vectors is input into an energy consumption semantic generation mechanism to generate an initial virtual energy consumption semantic sample set;

[0077] The generation of the initial virtual energy consumption semantic sample set comprises: generating a generative artificial intelligence inference model, the generative artificial intelligence inference model is used to establish a generative function of virtual energy consumption samples under semantic condition constraints, the generative function takes the set of time-continuous semantic state vectors as a conditional input and takes an energy consumption feature distribution as an output, through joint constraints of the environmental semantic component, the building use semantic component, and the personnel activity semantic component in the semantic state, conditional decoding of the energy consumption features in a semantic latent space is realized, and the initial virtual energy consumption semantic sample set is generated;

[0078] Distribution consistency correction is performed on the initial virtual energy consumption semantic sample set, a difference degree of the virtual energy consumption semantic sample set and a multi-source energy consumption data set in statistical features is calculated, internal parameters of the generative function are adjusted according to differences in energy consumption mean, variance, and distribution skewness, when the difference degree converges to a set threshold, a corrected virtual energy consumption semantic sample set is output, so that the distribution of the generated samples is consistent with the real distribution of the multi-source energy consumption data set;

[0079] performing time continuity correction on the corrected virtual energy consumption semantic sample set, outputting a time-continuous virtual energy consumption semantic sample set, the time continuity correction comprising: calculating semantic state change amplitudes of adjacent time steps, determining smoothing interpolation weights according to the change amplitudes, and performing interpolation and feature alignment operations on the virtual energy consumption semantic sample set in the time dimension, so that the generated samples maintain semantic continuity and time sequence stability in the time sequence;

[0080] performing semantic-level fusion on the time-continuous virtual energy consumption semantic sample set and the multi-source energy consumption dataset, performing feature mapping and weighted superposition in the same time step or adjacent time steps, and dynamically allocating fusion weights according to sample confidence, to obtain a fused energy consumption semantic sample set, the fused energy consumption semantic sample set containing both real observation features of the multi-source energy consumption dataset and semantic extension features of the virtual energy consumption semantic sample set.

[0081] In the embodiment, the generation of the energy consumption state response sequence comprises:

[0082] inputting the fused energy consumption semantic sample set into an improved liquid neural network, initializing an input layer of the liquid neural network according to feature vectors of each time step in the fused energy consumption semantic sample set, and setting an initial neural state as an energy consumption state response value of a previous time step, to establish an energy consumption dynamic equation for describing a continuous evolution relationship of a state of the liquid neural unit in the time dimension;

[0083] introducing a semantic driving regulation term into the energy consumption dynamic equation of the liquid neural network, and outputting a neural state sequence after semantic driving regulation, the semantic driving regulation term being used to adjust an activation output of the liquid unit according to semantic features in the fused energy consumption semantic sample set;

[0084] The generation of the neural state sequence after semantic driving regulation comprises: inputting a semantic feature vector into a regulation port of the liquid unit, multiplying an activation value of the liquid unit by a corresponding semantic weighting coefficient, calculating a semantic regulation gain according to a semantic feature change rate, increasing a unit response amplitude when the semantic change is sharp, and reducing a unit activation intensity when the semantic change is gentle; through the joint action of the semantic weighting and the regulation gain, the state update rate of the liquid unit is synchronized with the semantic change, and the neural state sequence after semantic driving regulation is output as an input of a time constant update mechanism;

[0085] The adaptive time constant updating mechanism is performed on the neural state sequence, and a time constant corrected neural state sequence is output, the adaptive time constant updating mechanism includes calculating the adjacent time step neural state change rate, combining the semantic feature change to construct a time adjustment coefficient, and real-time correcting the time constant of the liquid unit according to the time adjustment coefficient, shortening the time constant to enhance the response speed when the semantic change is violent, and lengthening the time constant to stabilize the output when the semantic change is gentle, and the time constant corrected neural state sequence is output;

[0086] The neural state sequence is input into the network output layer, and an energy consumption state response sequence is obtained after normalization processing, the energy consumption state response sequence represents the continuous time response law of energy consumption with semantic state change.

[0087] In the embodiment, the generation of the preliminary energy consumption prediction result includes:

[0088] An energy consumption time sequence prediction function is established based on the energy consumption state response sequence output by the improved liquid neural network, the energy consumption time sequence prediction function takes the energy consumption state response sequence as input and the energy consumption change amount of the future period as output, and forms a time-dependent mapping relationship between the input and the prediction output by setting a time step extension window;

[0089] The dependence relationship between each time step in the energy consumption state response sequence is calculated, and an energy consumption time sequence feature matrix is extracted, the feature extraction process includes: calculating the energy consumption state change rate and amplitude between adjacent time steps, determining the time correlation weight according to the change rate and amplitude, performing weighted operation on the input sequence according to the time correlation weight, generating the energy consumption time sequence feature matrix, and outputting the energy consumption time sequence feature matrix, which represents the evolution mode of energy consumption in the time dimension;

[0090] The energy consumption time sequence feature matrix is input into the energy consumption time sequence prediction function to perform prediction reasoning, and an energy consumption change trend sequence is output;

[0091] The prediction reasoning realizes the time extension of the energy consumption state change by establishing a nonlinear prediction mapping relationship in the time sequence space, each time step feature vector in the energy consumption time sequence feature matrix is input into the prediction function in time sequence, the prediction function jointly describes the dynamic dependence of the input feature and the future energy consumption change according to the continuous time weight coefficient and the nonlinear activation term, the error between the prediction output of each time step and the real energy consumption value is calculated in the prediction process, and the continuous time weight coefficient and the nonlinear activation parameter are dynamically adjusted according to the back propagation direction of the error, so as to minimize the objective constraint function of the prediction error square sum, when the prediction error converges to a set threshold, the optimized prediction parameters are used to calculate the energy consumption prediction values of multiple future time steps, and the energy consumption change trend sequence is formed through time sequence accumulation operation, and the energy consumption change trend sequence is output;

[0092] According to the energy consumption change trend sequence, the energy consumption change amplitude and direction of the future period are calculated, and a preliminary energy consumption prediction result is generated, which represents the energy consumption change trend of the campus air conditioning system in the future period.

[0093] In the embodiment, the generation of the combined self-adaptive optimized model parameter set comprises:

[0094] The preliminary energy consumption prediction result is time-synchronized with the real-time monitoring data in the multi-source energy consumption data set, and the predicted value and the measured energy consumption value at the same time step are differentially calculated to obtain a prediction error value, which is used to represent the deviation between the energy consumption state response sequence output by the improved liquid neural network and the Internet of Things measured energy consumption data;

[0095] The prediction error change rate is calculated, which is obtained by calculating the difference between the prediction error values at adjacent time steps and the ratio of the time interval, and is used to represent the speed of the prediction error changing with time, and the error change rate is used to judge the current energy consumption scene change trend, when the error change rate is greater than a set threshold, it is determined as an energy consumption mutation scene, and when the error change rate is lower than the threshold, it is determined as an energy consumption smooth scene;

[0096] The time constant of the improved liquid neural network is adjusted according to the prediction error change rate, so that the time constant is inversely proportional to the error change rate, the time constant is shortened to improve the response speed when the energy consumption mutation scene appears, and the time constant is lengthened to maintain the network output stable when the energy consumption smooth scene appears, and the neural state parameter after the time constant adjustment is output;

[0097] A joint self-adaptive optimization mechanism is established based on the prediction error value and the prediction error change rate, the joint self-adaptive optimization mechanism takes the improved liquid neural network and the energy consumption semantic generation mechanism as parallel optimization objects, and performs gradient update on the weight matrix parameters of the improved liquid neural network according to the prediction error square sum minimization objective function, the gradient update comprises: calculating the partial derivative of the loss function with respect to each parameter in the weight matrix, determining the gradient direction and correcting the weight value along the gradient descent direction, so that the updated weight value is equal to the original weight value minus the product of the learning rate and the loss function gradient, thereby gradually reducing the prediction error square sum, so that the energy consumption state response sequence generated by the neural state parameter after the time constant adjustment gradually approaches the real-time monitoring energy consumption data, and the prediction error information is input into the energy consumption semantic generation mechanism, the semantic generation parameters are corrected according to the prediction error distribution characteristics, so that the distribution of the generated virtual energy consumption semantic samples is consistent with the distribution of the real-time monitoring energy consumption data;

[0098] The improved liquid state neural network time constant and weight matrix parameters optimized by joint adaptive optimization are integrated with the energy consumption semantic generation mechanism semantic generation parameters after correction to form a model parameter set optimized by joint adaptive optimization, the model parameter set including the improved liquid state neural network and the energy consumption semantic generation mechanism optimized, and the improved liquid state neural network realizing parameter self-learning optimization through time constant dynamic adjustment and weight matrix gradient update.

[0099] In the embodiment, the generation of the semantic enhanced energy consumption prediction result includes:

[0100] The model parameter set optimized by joint adaptive optimization is called, the model parameter set including the improved liquid state neural network parameters after time constant dynamic adjustment and weight matrix gradient update and the energy consumption semantic generation mechanism semantic generation parameters after error distribution correction, and the model parameter set is loaded into the improved liquid state neural network and the energy consumption semantic generation mechanism to restore the optimized model state.

[0101] The energy consumption state response sequence output by the improved liquid state neural network is input into the energy consumption semantic generation mechanism, a semantic re-reasoning operation is performed under the constraint of the corrected semantic generation parameters, a semantic enhanced energy consumption prediction result is generated based on the semantic conditions, and the semantic enhanced energy consumption prediction result is used to reflect the semantic association characteristics of the energy consumption state and the environment, personnel activities and equipment operation conditions.

[0102] The semantic enhanced energy consumption prediction result is compared with the real-time monitoring data in the multi-source energy consumption data set, a new prediction error value is calculated according to the difference between the prediction result and the real-time monitoring data, and the time constant and weight matrix parameters of the improved liquid state neural network and the semantic generation parameters of the energy consumption semantic generation mechanism are iteratively updated online according to the prediction error value to form a dynamically updated model parameter set.

[0103] In the embodiment, the generation of the energy consumption trend curve and the load change report includes:

[0104] The semantic enhanced energy consumption prediction result and the dynamically updated model parameter set are called, the dynamically updated model parameter set including the improved liquid state neural network parameters and the energy consumption semantic generation mechanism semantic generation parameters after online iterative update.

[0105] The final energy consumption prediction result is data-arranged and time-series-unfolded, and an energy consumption prediction sequence of a continuous time period is generated according to the energy consumption values of each time step in the prediction result.

[0106] The energy consumption change rate per unit time and the load fluctuation amplitude are calculated according to the energy consumption prediction sequence, and an energy consumption trend curve is drawn, the energy consumption trend curve being used to reflect the energy consumption change trend of the air conditioning system of the campus building group in different time periods.

[0107] statistical analysis is performed on the energy consumption trend curve to extract an energy consumption peak period, a load fluctuation frequency and a change stability index, and a load change report is generated based on the same;

[0108] The generation of the load change report includes: performing time series expansion on the final energy consumption prediction result to extract the predicted energy consumption value of each time step; calculating the energy consumption difference and the change rate of adjacent time points to identify the load peak and trough interval; calculating the load fluctuation frequency and the energy consumption stability index based on the standard deviation of the change rate; calculating the prediction error rate in combination with the real-time monitoring energy consumption data; and finally generating the load change report according to the fluctuation characteristics, error distribution and stability index;

[0109] The energy consumption trend curve and the load change report are visualized and output, and the energy consumption prediction curve, the load fluctuation interval and the report abstract are displayed based on the time axis to form a visual display output of the final energy consumption prediction result.

[0110] Embodiment 1:

[0111] In order to verify the feasibility of the present application in implementation, the present application is applied to the air conditioning energy consumption prediction and dynamic optimization control scene of the teaching building group in the south area of a certain university. The area is composed of multiple teaching buildings, experimental buildings and office buildings, the internal air conditioning system structure is complex, including multi-split system, central chilled water chiller and split air conditioning terminal equipment, and the equipment operation is affected by many factors such as weather change, course arrangement and personnel density fluctuation. When facing such complex multi-source energy consumption characteristics, the traditional energy consumption prediction method often fails to capture the implicit relationship between personnel activities and environmental factors, especially when there is missing, uneven distribution or different sampling period in the energy consumption monitoring data, the stability and accuracy of the prediction model decrease significantly, which cannot meet the demand of real-time energy saving regulation. In view of this problem, the campus air conditioning energy consumption prediction method based on AIGC reasoning proposed by the present application is introduced into the energy consumption management system of the building group to verify its intelligent prediction and dynamic adaptation ability under the condition of incomplete multi-source data and severe environmental change.

[0112] In the implementation process, the air conditioning system operation parameters, personnel density information, environmental temperature and humidity and external meteorological conditions and other multi-source environmental data of each building in the teaching building group are collected in real time through the Internet of Things collection system, and the obtained raw data is subjected to time synchronization and abnormality elimination processing. For the sensor data with missing values, interpolation completion and smoothing processing technology is used for repair, so that the energy consumption data maintains continuity in time dimension. Subsequently, the data standardization module is used to normalize variables of different dimensions to form a unified structure of multi-source energy consumption dataset. The dataset not only contains the real-time operation characteristics of air conditioning equipment, but also contains the change information of environmental and personnel behavior patterns, which provides complete input conditions for subsequent semantic modeling and reasoning.

[0113] After the data processing is completed, semantic feature extraction is performed on the multi-source data by using a feature encoding network. The feature encoding network captures the change law of energy consumption features at short-term and long-term time scales through multi-scale convolution and recursive structure, and establishes the associated representation between building use semantics, personnel activity semantics and environment semantics in the semantic embedding layer. After semantic mapping and time continuity correction, a set of semantic state vectors that can represent the dynamic characteristics of building energy consumption is formed. These semantic state vectors reflect the dynamic change trend of building energy consumption affected by multiple factors in different time periods.

[0114] On this basis, an energy consumption semantic generation mechanism is called, which takes the set of semantic state vectors as input and generates virtual energy consumption semantic samples under semantic constraints through a generative inference process. This mechanism can perform semantic consistency completion on the missing data in the semantic latent space, thereby solving the problem of insufficient samples caused by data sparsity. The generated virtual energy consumption semantic samples are fused with the multi-source energy consumption data set at the semantic level to construct a fused energy consumption semantic sample set, making the distribution of training samples more complete and balanced. This significantly improves the prediction stability of the model under abnormal weather or special use scenarios, enabling it to have adaptive generalization ability when the data conditions change.

[0115] The fused semantic sample set is input into an improved liquid neural network. This network introduces a semantic-driven regulation term and an adaptive time constant updating mechanism based on the traditional liquid unit, enabling the network to dynamically adjust the evolution rate of neural states according to semantic changes. When external environmental fluctuations or sudden changes in personnel density cause sharp changes in energy consumption features, the semantic-driven regulation term will amplify the neural response amplitude to quickly capture energy consumption mutations. In the stable stage of energy consumption, the adaptive time constant mechanism prolongs the state update interval to maintain output stability. Through this continuous time modeling method, the energy consumption state response sequence can truly reflect the dynamic law of building energy consumption changes with the environment and semantics.

[0116] Subsequently, an energy consumption time series prediction function is established using the energy consumption state response sequence to predict future energy consumption trends through time extension and nonlinear mapping inference. The prediction result is then differenced with the real-time collected energy consumption monitoring data to obtain the prediction error information. To ensure the long-term stability and high precision of the model in operation, the invention further introduces a joint adaptive optimization mechanism, which forms a closed-loop feedback relationship between the liquid neural network and the energy consumption semantic generation mechanism. When the prediction error increases, the system automatically performs gradient update and semantic parameter correction to realize joint optimization of the weight matrix and semantic generation parameters, enabling the model parameters to continuously self-adjust and learn with changes in the scene.

[0117] In the continuous running phase, the optimized improved liquid state neural network outputs the energy consumption state response sequence, which is input into the energy consumption semantic generation mechanism again to perform semantic re-reasoning operation, and a semantic enhanced energy consumption prediction result is generated. The semantic enhanced result not only has accurate prediction ability for future energy consumption change, but also can reveal the semantic cause and effect relationship behind the energy consumption change, such as 'outdoor temperature rises sharply to cause cold load increase' or 'personnel density change causes energy consumption fluctuation'. Finally, the system visualizes the energy consumption prediction result to output the energy consumption trend curve and load change report in the form of time sequence, providing decision reference for energy consumption management department. The energy consumption trend curve directly shows the trend of energy consumption change in each period, and the load change report comprehensively reflects the energy consumption fluctuation frequency, peak period and stability index, providing basis for energy saving regulation and control.

[0118] Through the above implementation process, the semantic understanding, dynamic reasoning and adaptive prediction of the campus air conditioning system energy consumption are realized, and the shortcomings of the traditional model in data missing, scene change and prediction lag are overcome. The method not only improves the prediction accuracy and timeliness, but also enhances the adaptive learning ability of the model in complex environment, providing a new technical path for intelligent energy saving management of college building group.

[0119] Table 1 Performance comparison of campus air conditioning energy consumption prediction method based on AIGC reasoning

[0120] Method name Average prediction error (kWh) Maximum error (kWh) Prediction stability coefficient (%) Accuracy retention rate under data missing condition (%) Prediction delay time (seconds) Traditional LSTM model 5.82 12.64 86.7 71.3 4.5 GRU neural network 5.43 11.89 88.1 73.5 4.2 Transformer prediction framework 4.97 10.72 90.3 77.6 3.9 Energy consumption prediction method based on digital twinning 4.36 9.85 91.8 82.4 3.7 The method of the present invention 2.91 6.38 96.2 91.7 2.4

[0121] As can be seen from Table 1, the campus air conditioning energy consumption prediction method based on AIGC reasoning proposed by the present application is significantly better than the traditional prediction method in various indicators. In terms of average prediction error, the method of the present application is only 2.91kWh, which is about 50% lower than the traditional LSTM model, showing the significant improvement of the method in energy consumption prediction accuracy. In terms of maximum error, the prediction result of the method of the present application is more stable, with error peak controlled within 6.38kWh, indicating that the model can still maintain high prediction consistency when facing sudden environmental changes.

[0122] In addition, the prediction stability coefficient of the method of the present application reaches 96.2%, which is about 6 percentage points higher than that of the Transformer prediction framework, indicating that the AIGC semantic generation mechanism can effectively enhance the semantic understanding ability of the model to complex energy consumption characteristics, so that the model can maintain output stability in long time running. It is worth noting that under the condition of data missing, the accuracy retention rate of the method of the present application can still reach 91.7%, which is more than 20% higher than that of the traditional LSTM method, which reflects the robustness and generalization ability of the present application under the condition of sample sparsity or data anomaly.

[0123] In terms of prediction delay, the response time of the method of the application is 2.4 seconds, which is significantly lower than the average level of other models, indicating that the improved liquid neural network can realize real-time energy consumption trend prediction and rapid response under the support of continuous time dynamic modeling and parameter adaptive updating mechanism. The application realizes the comprehensive improvement of prediction accuracy, stability and real-time performance by introducing semantic driving regulation term, adaptive time constant updating mechanism and AIGC generation reasoning completion strategy, and provides efficient and reliable technical support for dynamic energy consumption optimization of the intelligent campus air conditioning system.

[0124] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A campus air conditioning energy consumption prediction method based on AIGC inference, characterized in that, Includes the following steps: Collect multi-source environmental data of air conditioning systems in campus buildings, and perform time synchronization, anomaly removal, interpolation completion and normalization on the data to form a structured multi-source energy consumption dataset. Extract key variables and use a feature encoding network to generate a semantic state vector set; The semantic state vector set is input into the energy consumption semantic generation mechanism to generate virtual energy consumption semantic samples that are consistent with the real distribution, thus forming a fused energy consumption semantic sample set. The fused energy consumption semantic sample set is input into the improved liquid neural network to construct a continuous-time energy consumption dynamic equation and output an energy consumption state response sequence. Based on the energy consumption state response sequence, an energy consumption time series prediction function is established to predict the energy consumption change trend in future periods and generate preliminary energy consumption prediction results. The preliminary energy consumption prediction results are compared with the real-time monitoring data in the multi-source energy consumption dataset, the prediction error value is calculated, the joint adaptive optimization action is performed, and the set of model parameters after joint adaptive optimization is output. The optimized energy consumption state response sequence is input into the energy consumption semantic generation mechanism, and a semantic re-inference operation is performed under the constraints of the corrected semantic generation parameters to generate semantically enhanced energy consumption prediction results. The final energy consumption forecast results are visualized and output, generating energy consumption trend curves and load change reports.

2. The campus air conditioning energy consumption prediction method based on AIGC inference according to claim 1, characterized in that, The generation of the semantic state vector set includes: Time synchronization, anomaly removal, and missing value completion are performed on the raw sensor data in the multi-source energy consumption dataset. Data from different sources are aligned on a unified time scale, and data gaps are repaired by linear interpolation and median smoothing methods to obtain a time-continuous multi-source feature sequence. The multi-source feature sequence is input into the feature standardization module to perform normalization processing on numerical features of different dimensions, and to construct a multi-dimensional input feature matrix. The feature matrix is ​​input into a feature encoding network with multi-scale convolution and recursion structure. Short-term energy consumption change features are extracted through local convolutional units, and long-term change trends are captured through time recursive units, generating an intermediate feature representation sequence that describes the dynamic change pattern of energy consumption. Semantic mapping is performed on the intermediate feature representation sequence. Feature importance weights are calculated based on the contribution of each input feature to energy consumption change. Weighted fusion is performed based on the semantic correlation between features and energy consumption. At each time step, all features are weighted and summed to obtain the semantic embedding representation. The semantic embedding representation is input to the temporal fusion unit for temporal continuity correction. Smooth interpolation and dynamic adjustment are performed based on the semantic change magnitude at adjacent time points, and a temporally continuous semantic state vector set is output.

3. The campus air conditioning energy consumption prediction method based on AIGC inference according to claim 1, characterized in that, The formation of the fused energy consumption semantic sample set includes: Input the semantic state vector set into the energy consumption semantic generation mechanism to generate an initial virtual energy consumption semantic sample set; Perform distribution consistency correction on the initial virtual energy consumption semantic sample set, calculate the difference in statistical characteristics between the virtual energy consumption semantic sample set and the multi-source energy consumption dataset, and output the corrected virtual energy consumption semantic sample set when the difference converges to a set threshold. Perform time continuity correction on the corrected virtual energy consumption semantic sample set, and output a time-continuous virtual energy consumption semantic sample set; The time-continuous virtual energy consumption semantic sample set is semantically fused with the multi-source energy consumption dataset. Feature mapping and weighted superposition are performed in the same or adjacent time steps. The fusion weights are dynamically allocated according to the sample confidence to obtain the fused energy consumption semantic sample set.

4. The campus air conditioning energy consumption prediction method based on AIGC inference according to claim 1, characterized in that, The generation of the energy consumption state response sequence includes: The fused energy consumption semantic sample set is input into the improved liquid neural network. The input layer of the liquid neural network is initialized according to the feature vector of each time step in the fused energy consumption semantic sample set, and a continuous-time energy consumption dynamic equation is established. A semantically driven regulation term is introduced into the energy consumption dynamic equation of the liquid neural network, and the neural state sequence after semantically driven regulation is output. An adaptive time constant update mechanism is applied to the neural state sequence to output a neural state sequence corrected by the time constant. The neural state sequence is input into the network output layer and normalized to obtain the energy consumption state response sequence.

5. The campus air conditioning energy consumption prediction method based on AIGC inference according to claim 1, characterized in that, The generation of the preliminary energy consumption prediction results includes: An energy consumption time series prediction function is established based on the energy consumption state response sequence output by the improved liquid neural network. A time-dependent mapping relationship between the input and the predicted output is formed by setting a time step extension window. Calculate the dependencies between time steps in the energy consumption state response sequence and extract the energy consumption time series feature matrix, and output the energy consumption time series feature matrix; The energy consumption time series feature matrix is ​​input into the energy consumption time series prediction function to perform prediction inference and output the energy consumption change trend sequence. Based on the energy consumption change trend sequence, calculate the magnitude and direction of energy consumption changes in future periods to generate preliminary energy consumption prediction results.

6. The campus air conditioning energy consumption prediction method based on AIGC inference according to claim 1, characterized in that, The generation of the model parameter set after joint adaptive optimization includes: The preliminary energy consumption prediction results are synchronized with the real-time monitoring data in the multi-source energy consumption dataset. The difference between the predicted value and the measured energy consumption value at the same time step is calculated to obtain the prediction error value. Calculate the prediction error change rate, and judge the current energy consumption scenario change trend based on the error change rate. When the error change rate is greater than the set threshold, it is determined to be an energy consumption sudden change scenario. When the error change rate is less than the threshold, it is determined to be an energy consumption stable scenario. The time constant of the liquid neural network is adjusted and improved based on the rate of change of the prediction error, so that the time constant is inversely proportional to the rate of change of the error. When a sudden energy consumption scenario occurs, the time constant is shortened to improve the response speed, and when a stable energy consumption scenario occurs, the time constant is extended to maintain the stability of the network output. The output is the neural state parameters adjusted by the time constant. A joint adaptive optimization mechanism is established based on the prediction error value and the prediction error change rate. According to the objective function of minimizing the sum of squared prediction errors, the gradient update is performed on the weight matrix parameters of the improved liquid neural network. At the same time, the prediction error information is input into the energy consumption semantic generation mechanism, and its semantic generation parameters are corrected according to the distribution characteristics of the prediction error. The improved liquid neural network time constant and weight matrix parameters, which have undergone joint adaptive optimization, are integrated with the modified energy consumption semantic generation mechanism semantic generation parameters to form a set of model parameters after joint adaptive optimization.

7. A campus air conditioning energy consumption prediction method based on AIGC inference according to claim 1, characterized in that, The generation of the semantically enhanced energy consumption prediction results includes: The jointly adaptively optimized model parameter set is invoked and loaded into the improved liquid neural network and energy consumption semantic generation mechanism; The energy consumption state response sequence output by the improved liquid neural network is input into the energy consumption semantic generation mechanism. Under the constraints of the modified semantic generation parameters, a semantic re-inference operation is performed to generate semantically enhanced energy consumption prediction results based on semantic conditions. The semantically enhanced energy consumption prediction results are compared with the real-time monitoring data in the multi-source energy consumption dataset. A new prediction error value is calculated based on the difference between the prediction results and the real-time monitoring data. Based on the prediction error value, the time constant and weight matrix parameters of the improved liquid neural network and the semantic generation parameters of the energy consumption semantic generation mechanism are updated online to form a dynamically updated set of model parameters.

8. A campus air conditioning energy consumption prediction method based on AIGC inference according to claim 1, characterized in that, The generation of the energy consumption trend curve and load change report includes: Call the semantically enhanced energy consumption prediction results and the dynamically updated set of model parameters; The final energy consumption prediction results are processed and expanded into a time series. Based on the energy consumption values ​​at each time step in the prediction results, a continuous time period energy consumption prediction sequence is generated. Calculate the rate of change of energy consumption per unit time and the load fluctuation amplitude based on the energy consumption prediction sequence, and plot the energy consumption trend curve. Perform statistical analysis on the energy consumption trend curve, extract peak energy consumption periods, load fluctuation frequency and change stability indicators, and generate load change reports accordingly; The energy consumption trend curve and load change report are visualized and output, and the energy consumption prediction curve, load fluctuation range and report summary are displayed on the time axis to form a visual display of the final energy consumption prediction results.

Images