Energy management method and system, electronic equipment and storage medium

Abstract

The invention relates to the technical field of Internet of Things, and provides an energy management method and system, electronic equipment and a storage medium, and the method comprises the steps: obtaining symbol content and perception content corresponding to electrical equipment; based on a sharing layer of an energy management model, performing semantic feature extraction on the symbol content and the perception content, and fusing the extracted symbol semantic features and perception semantic features to obtain fused semantic features; and based on the task head of the energy management model, applying the fused semantic features to generate a corresponding equipment control strategy, the equipment control strategy being used for adjusting the operation state of the electrical equipment. According to the method, through multi-source heterogeneous data fusion, semantic-level feature modeling and dynamic task adaptation, intelligent decision-making of energy management is realized, the limitation that traditional energy management depends on manual rules or single data sources is broken through, and the accuracy, adaptability and individuation level of energy management are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things, and in particular to an energy management method, system, electronic device and storage medium. Background Art

[0002] Against the backdrop of improved living standards and the popularization of home appliances, household electricity consumption has increased and the peak-to-valley difference has put pressure on power supply facilities, while the problem of household energy waste has become prominent. In this situation, home energy management brought about by smart homes has become a key way to optimize energy distribution, improve energy efficiency and reduce expenses.

[0003] However, existing home energy management solutions have significant flaws. Traditional automated management solutions often rely on manually pre-set rules and lack the ability to adapt to user behavior and the dynamic changes in home energy and grid supply and demand. Furthermore, relying solely on local sensor data collection results in a single information source, which affects decision-making accuracy. Summary of the Invention

[0004] The present invention provides an energy management method, system, electronic device and storage medium to address the defects of related technologies such as poor flexibility and poor decision-making accuracy caused by reliance on manual rules or a single data source for energy management.

[0005] The present invention provides an energy management method, comprising: Acquiring symbolic content and perceptual content corresponding to an electrical device, wherein the symbolic content includes external factors and / or user habits that affect the use of the electrical device, and the perceptual content includes device operation information and / or environmental perception information of the electrical device; Based on the shared layer of the energy management model, semantic features of the symbol content and the perception content are extracted respectively, and the extracted symbol semantic features and perception semantic features are fused to obtain fused semantic features corresponding to the electrical device; Based on the task head of the energy management model, the fused semantic features are applied to generate a corresponding device control strategy, where the device control strategy is used to adjust the operating state of the electrical device.

[0006] According to an energy management method provided by the present invention, the shared layer based on the energy management model extracts semantic features from the symbolic content and the perceptual content respectively, and fuses the extracted symbolic semantic features and perceptual semantic features to obtain a fused semantic feature corresponding to the electrical device, including: Based on the symbol content processing module of the shared layer, semantic feature extraction is performed on the symbol content to obtain the symbol semantic feature; Extracting semantic features of the perceived content based on the perceived content processing module of the shared layer to obtain the perceived semantic features; Based on the fusion processing module of the shared layer, the symbolic semantic feature and the perceptual semantic feature are fused to obtain the fused semantic feature.

[0007] According to an energy management method provided by the present invention, the symbol content processing module based on the shared layer extracts semantic features from the symbol content to obtain the symbol semantic features, including: Based on the symbol content processing module, when a change in the symbol content is detected, the attention mechanism is used to extract semantic features of the symbol content to obtain the symbol semantic features.

[0008] According to an energy management method provided by the present invention, the task head includes a memory head and a control head. The task head based on the energy management model applies the fused semantic features to generate a corresponding device control strategy, including: Based on the memory head, applying the fused semantic features to identify user habits according to a preset frequency period, converting the identified user habits into a user habit symbol sequence, and updating the pre-recorded user habit symbol sequence when the user habit symbol sequence changes; Based on the control head, the fused semantic features are applied to generate corresponding device control strategies.

[0009] According to an energy management method provided by the present invention, the control head includes a real-time task control head and a planned task control head, and the device control strategy includes an immediate task event and a planned task event sequence; Accordingly, the generating of a corresponding device control strategy based on the control head and applying the fused semantic features includes: Based on the real-time task class control header, applying the fused semantic features, predicting the expected state of the electrical device, and generating an immediate task event according to the expected state of the electrical device and the actual state of the electrical device; Based on the planned task class control header, the fused semantic features are applied to generate an expected planned task event sequence, where the planned task event sequence includes the execution order of each task event and the corresponding time nodes.

[0010] According to an energy management method provided by the present invention, the energy management model is obtained based on centralized pre-training before deployment and fine-tuning after deployment. The pre-training step of the energy management model includes: Based on the training data set, a multi-task learning method is used to pre-train each module of the shared layer in the energy management model to obtain a pre-trained shared layer; Based on the specific tasks of each task head in the energy management model, each task head is pre-trained, and a multi-task learning method is adopted to parallelly optimize the loss functions of different tasks on the pre-trained shared layer to synchronously adjust the pre-trained shared layer to obtain a pre-trained energy management model.

[0011] According to an energy management method provided by the present invention, the fine-tuning of the energy management model is implemented based on online incremental learning after the model is deployed. The fine-tuning of the energy management model includes fine-tuning the network parameters of the fusion processing module and each control head in the energy management model, and does not fine-tune the network parameters of the symbolic content processing module, the perception content processing module and the memory head in the energy management model.

[0012] According to an energy management method provided by the present invention, the energy management model is deployed on a local distributed node, and the user habits in the symbolic content and the perception content are generated and acquired locally.

[0013] The present invention also provides an energy management system, comprising: a content acquisition module, configured to acquire symbolic content and perceptual content corresponding to an electrical device, wherein the symbolic content includes external factors and / or user habits that affect the use of the electrical device, and the perceptual content includes device operation information and / or environmental perception information of the electrical device; a feature extraction module for extracting semantic features of the symbolic content and the perceptual content based on the shared layer of the energy management model, and fusing the extracted symbolic semantic features and perceptual semantic features to obtain a fused semantic feature corresponding to the electrical device; A strategy generation module is used to generate a corresponding device control strategy based on the task header of the energy management model and the application of the fusion semantic features, wherein the device control strategy is used to adjust the operating state of the electrical device.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements any of the above-mentioned energy management methods when executing the computer program.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which implements any of the above-mentioned energy management methods when executed by a processor.

[0016] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above energy management methods.

[0017] The energy management method, system, electronic device, and storage medium provided by the present invention acquire the symbolic content and sensory content corresponding to electrical devices. The symbolic content integrates external factors and / or user habits that influence appliance usage, providing a macro-level decision-making basis for energy management. The sensory content includes real-time collected device operation information and environmental perception information, ensuring that decisions are close to actual scenarios. Semantic features are extracted from the symbolic and sensory content through the shared layer of the energy management model, capturing the deep semantic connections between different information sources and improving the ability to understand complex scenarios. Semantic feature extraction and fusion transform discrete multi-source information into structured features, avoiding the limitations of a single information source and achieving global optimization of energy management. Furthermore, the organic integration of symbolic logic and sensory data allows decisions to take into account user habits and real-time needs, reducing misjudgments and improving the accuracy, adaptability, and personalization of energy management. Furthermore, by generating device control strategies based on fused semantic features, the model can respond to changes in external factors or environmental conditions in real time, dynamically optimizing control strategies and avoiding static, "one-size-fits-all" rules, thereby improving energy efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or related technologies, the following is a brief introduction to the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 This is one of the flow charts of the energy management method provided by the present invention; Figure 2 This is the second flow chart of the energy management method provided by the present invention; Figure 3 This is the third flow chart of the energy management method provided by the present invention; Figure 4 It is a structural diagram of the energy management model provided by the present invention; Figure 5 This is a schematic diagram of the overall process of the home energy management method based on the distributed AI model provided by the present invention; Figure 6 It is a structural diagram of the energy management system provided by the present invention; Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0021] With socioeconomic development and improvements in people's quality of life, the penetration and frequency of household appliances continue to rise, and their power consumption has become a core component of household energy consumption. At the same time, the peaks and valleys in electricity consumption by household appliances place high demands and challenges on the carrying capacity and stability of the power supply infrastructure. However, current household energy efficiency is generally low, with significant energy waste, which undoubtedly conflicts sharply with the growing demand for energy conservation and consumption reduction. Against this backdrop, the rise of smart home technology has opened up new avenues for home energy management. As a key application area in the smart home sector, home energy management achieves optimal energy allocation and efficient utilization by precisely controlling the operating parameters and startup timing of various electrical devices, becoming a key means of improving household energy efficiency and reducing energy costs.

[0022] Existing smart home technologies offer preliminary solutions for home energy management through real-time energy consumption monitoring and automated control. However, these solutions still face numerous challenges in practical application. One type of solution uses traditional automation methods, which rely heavily on experienced technicians to pre-set rules. These systems operate strictly according to established logic and lack the ability to deeply adapt to individual user needs and behavior patterns. This inability to flexibly adjust to the evolving needs of household appliances, the dynamic changes in household energy demand, and the real-time fluctuations in grid power supply makes it difficult to achieve precise and efficient energy management.

[0023] Another category of solutions uses artificial intelligence (AI) to address personalization and adaptability needs. However, these solutions often operate offline, lacking timely access to algorithmic iterations and upgrades, making it difficult to fully absorb cutting-edge technological advances. Furthermore, relying solely on local sensor data collection and a single source of information limits decision-making, reducing the scientific nature and accuracy of energy management. Furthermore, some AI solutions leverage the powerful computing power of the cloud to perform tasks, but these solutions require large amounts of data to be transmitted over the internet, which undoubtedly increases the risk of private data leakage, seriously threatening user information security, and lacking reliability.

[0024] To address this issue, the present invention provides an energy management method based on a distributed AI model to address the issues of personalization, adaptability, flexibility, reliability, and user privacy and security in smart home energy management. The technical solution provided by the present invention is described in detail below.

[0025] Figure 1 This is one of the flow charts of the energy management method provided by the present invention, such as Figure 1 As shown, the method includes: Step 110 , obtaining symbol content and perception content corresponding to the electrical device, wherein the symbol content includes external factors and / or user habits that affect the use of the electrical device, and the perception content includes device operation information and / or environmental perception information of the electrical device.

[0026] Specifically, symbolic content refers to indirect observational data related to electrical equipment usage, expressed in symbolic form. This data includes information such as external factors and / or user habits. Symbolic content uses abstract symbols to describe macroeconomic conditions and user behavior patterns, providing logical rules and historical experience for energy management.

[0027] External factors here refer to public information obtained from external sources, such as weather forecasts, extreme weather warnings, power supply status, grid load status, and energy prices. This information can be downloaded from a network server and converted into a symbol sequence to ensure a uniform data format. External factors are public, timely, and macroscopic, and therefore constitute public data outside the scope of user privacy.

[0028] User habits refer to information derived from historical user behavior, such as sleep and rest schedules, activity ranges, and device associations. These can be recorded as symbol sequences through local logs or user configurations, forming a personalized dataset. User habits are private, long-term, and personalized, making them core data within the user privacy domain.

[0029] Perception content refers to physical quantities or status codes directly collected by sensors and devices that reflect real-time conditions. These data types include device operating information and / or environmental perception information. By monitoring device status and environmental parameters in real time, perception content provides dynamic data support for energy management, enabling immediate response and proactive planning.

[0030] Here, device operation information refers to data directly collected through built-in sensors or interfaces in electrical devices, such as power, temperature, operating status code, mode status code, and reservation requests. Environmental perception information refers to information collected by sensors deployed in the home environment (such as temperature and humidity sensors, biometric sensors, and cameras), such as temperature and humidity, light intensity, and the behavior of family members.

[0031] Step 120 , based on the shared layer of the energy management model, semantic features are extracted from the symbolic content and the perceptual content respectively, and the extracted symbolic semantic features and perceptual semantic features are fused to obtain fused semantic features corresponding to the electrical equipment.

[0032] It's important to note that an energy management model is an AI model based on machine learning or deep learning, used to optimize energy distribution and device control in homes or industrial settings. Its core goal is to generate efficient, personalized device control strategies by analyzing multi-source heterogeneous data (such as symbolic and perceptual content), achieving a balance between energy conservation and consumption reduction and user experience.

[0033] Specifically, the energy management model can include a shared layer, which is the core module of the energy management model and is responsible for extracting common semantic features from the symbolic content and perceptual content, providing a unified feature representation for subsequent task heads. Specifically, the shared layer can extract semantic features from the input symbolic content and perceptual content respectively, converting the abstract logic of the symbolic content and the real-time data of the perceptual content into semantic vectors that the model can understand.

[0034] It's understandable that symbolic content (such as weather forecasts and user schedules) is inherently discrete and structured data. The shared layer can use an attention mechanism to extract semantic information from the symbolic content to obtain symbolic semantic features. Here, symbolic semantic features refer to abstract semantic vectors extracted from the symbolic content, reflecting macroscopic conditions and user behavior logic. Perceptual content (such as temperature and device power) is continuous, unstructured numerical data. The shared layer can use convolutional neural networks (CNNs) to extract features and long short-term memory (LSTM) networks to handle temporal dependencies, thereby obtaining perceptual semantic features. Here, perceptual semantic features refer to temporal or spatial semantic vectors extracted from the perceptual content, reflecting the device and environmental states.

[0035] After extracting the symbolic semantic features and perceptual semantic features, they can be fused to generate a fused feature that combines macroscopic understanding with microscopic responsiveness. Here, when fusing the symbolic semantic features and perceptual semantic features, the two feature vectors can be directly concatenated to form a fused semantic feature. Alternatively, weights can be assigned to the symbolic and perceptual features, and fusion can be performed using a weighted summation approach. This is not specifically limited in the present embodiment. It should be understood that a fused semantic feature is a unified vector representation of the fusion of symbolic semantic features (i.e., logical rules) and perceptual semantic features (i.e., real-time status). It contains both symbolic logic (e.g., "High temperature requires cooling") and perceptual status (e.g., "Current temperature is 28°C"), dynamically reflecting the interaction between symbolic rules and real-time data, and providing more interpretable feature input for subsequent task heads.

[0036] Step 130 : Based on the task header of the energy management model, the fused semantic features are applied to generate a corresponding device control strategy, which is used to adjust the operating status of the electrical device.

[0037] It should be noted that the task head is the output layer module designed for specific functions or tasks in the energy management model. Its core function is to convert the fused semantic features extracted by the shared layer into specific equipment control strategies.

[0038] Specifically, after receiving the fused semantic features output by the shared layer, the task head can use neural network methods to process the features and semantics to generate corresponding device control strategies. For example, the fused semantic features can be used to predict the expected state of an electrical device and compare it with its actual state. When the actual state differs from the expected state, an immediate device control task event is generated, such as "Adjust the air conditioner temperature from 26°C to 22°C." For another example, by decoding the fused semantic features, a timed task event sequence can be generated, such as "Schedule the washing machine to start at 3:00 AM."

[0039] It's understandable that a device control strategy refers to a series of tasks that change the operating state of electrical devices, such as "Set the bedroom air conditioner to 26°C for dehumidification" or "Schedule the washing machine to start at 3:00 AM." A control strategy here is an event instruction, not a control signal. The energy management model only generates event instructions, while the control signals for the specific electrical devices required to implement these event instructions are handled by subsequent modules based on the specific device model and interface. For example, the control strategy "Set the bedroom air conditioner to 26°C for dehumidification" is actually an API (Application Programming Interface) call, transmitted to the IoT remote controller of a certain brand of air conditioner. The device then converts it into an actual control signal specific to that brand of air conditioner, controlling the mode and temperature of the bedroom air conditioner.

[0040] The method provided by an embodiment of the present invention obtains the symbolic content and perceptual content corresponding to electrical devices. The symbolic content integrates external factors and / or user habits that influence appliance usage, providing a macro-level decision-making basis for energy management. The perceptual content includes real-time collected device operating information and environmental perception information, ensuring that decisions are close to actual scenarios. Semantic features are extracted from the symbolic and perceptual content separately through the shared layer of the energy management model, capturing the deep semantic connections between different information sources and improving the ability to understand complex scenarios. Semantic feature extraction and fusion transform discrete multi-source information into structured features, not only avoiding the limitations of a single information source and achieving global optimization of energy management, but also organically integrating symbolic logic with perceptual data, enabling decisions to take into account user habits and real-time needs, reducing misjudgments and improving the accuracy, adaptability, and personalization of energy management. Furthermore, by generating device control strategies based on fused semantic features, the model can respond in real time to changes in external factors or environmental conditions, dynamically optimizing control strategies and avoiding static, "one-size-fits-all" rules, thereby improving energy efficiency.

[0041] Based on the above embodiments, Figure 2 This is the second flow chart of the energy management method provided by the present invention, such as Figure 2 As shown, the method includes: Step 110 , obtaining symbol content and perception content corresponding to the electrical device, wherein the symbol content includes external factors and / or user habits that affect the use of the electrical device, and the perception content includes device operation information and / or environmental perception information of the electrical device.

[0042] Specifically, the input of the energy management model includes two parts: symbolic content and perceptual content. The symbolic content is a converted symbol sequence, which includes two parts: one is the announcement information from external information sources, such as weather forecasts, extreme weather warnings, power supply status, grid load status, energy prices, etc. The announcement information is downloaded from the network in the form of a symbol sequence, and the conversion process is performed on the network server; the other is the locally recorded user habits, such as work and rest time, activity range, device association, etc. User habits are saved locally in the form of a symbol sequence.

[0043] Perception input is information acquired through sensors and electrical devices, such as sensor data, device operating status, family member behavior, and reservation requests and mode settings for large electrical devices. Perception input can be a fixed-size set of values, including data values ​​collected by numerical sensors, device operating status codes, family member behavior status codes collected by biometric sensors, and device reservation and mode status codes.

[0044] Step 120: Based on the shared layer of the energy management model, semantic features are extracted from the symbolic content and the perceptual content respectively, and the extracted symbolic semantic features and perceptual semantic features are fused to obtain a fused semantic feature corresponding to the electrical device; It should be noted that the function of the shared layer is to extract common features and semantic information from the input. These extracted features and semantics are represented by data within the neural network. The shared layer consists of three submodules: the symbolic content processing module, the perceptual content processing module, and the fusion processing module. The symbolic content processing module specifically processes the symbolic content of the input, while the perceptual content processing module specifically processes the perceptual content of the input. These two modules initially extract features and semantics of the corresponding content, which are then input into the fusion processing module. The fusion processing module combines the preliminary features and semantics generated by these two modules and processes them into features and semantics that can be used by the task head.

[0045] Specifically, the above step 120 specifically includes: Step 121 : Based on the symbol content processing module of the shared layer in the energy management model, semantic features of the symbol content are extracted to obtain symbol semantic features.

[0046] It is understandable that the symbol content processing module can use the attention mechanism to extract the semantic information of the symbol content to obtain the symbol semantic features. For example, the symbol content processing module can be implemented using a Transformer encoder network, which includes four layers: multi-head self-attention, residual sum normalization, feedforward fully connected layer, and residual sum normalization. For another example, the symbol content processing module can also be implemented using the encoder structure of models such as BERT (Bidirectional Encoder Representations from Transformer, based on Transformer bidirectional encoder representations), ALBERT (a lightweight variant of the BERT model), and XLNet (Generalized Autoregressive Pretraining for Language Understanding). The embodiment of the present invention does not impose specific restrictions on this.

[0047] Specifically, discrete symbol sequences can first be converted into dense vector representations. For example, external announcement information (i.e., external factors such as weather forecasts) can be converted into vectors using pre-trained word embeddings (e.g., Word2Vec, BERT). User habits (e.g., sleep schedules) can be mapped into vectors using numerical or categorical encoding. Then, the multi-head self-attention layer in the symbol content processing module calculates the correlation weights between input vectors, capturing long-range dependencies (e.g., the association between "extreme weather warnings" and "frequency of air conditioning use"). Semantic features from different subspaces are extracted in parallel using multiple attention heads. The outputs of these attention heads are concatenated and linearly transformed to generate features that enhance semantic associations. Next, a residual sum normalization layer (i.e., residual connection and layer normalization) preserves the original input information. This involves adding the input vector to the features output by the multi-head self-attention layer to mitigate gradient spuriousness. The resulting features are then normalized to stabilize training. Subsequently, a feed-forward fully connected layer applies a nonlinear transformation to the normalized features to enhance their expressive power. Finally, the features output by the feedforward fully connected layer undergo a second round of residual connections and layer normalization to stabilize the training process. After multi-layer Transformer encoding, the symbolic semantic features output by the symbol content processing module are obtained.

[0048] Furthermore, step 121 specifically includes: Based on the symbol content processing module, when a change in the symbol content is detected, the attention mechanism is used to extract the semantic features of the symbol content to obtain the symbol semantic features.

[0049] Specifically, since the symbol content mainly includes announcement data downloaded / pulled from the network, remote or cloud, as well as locally recorded user habits, and this data does not change frequently, the symbol content processing module can be set to use the attention mechanism to extract semantic features of the symbol content only when a change in the symbol content is detected. If it is detected that the symbol content has not changed, the symbol semantic features extracted from the previous extraction can be directly used for subsequent feature fusion.

[0050] It is understood that when detecting whether a symbol's content has changed, a lightweight detection method using continuous hashing or checksums can be used. For example, a hash value or checksum (such as MD5 or CRC32) can be generated for the symbol's content and periodically compared with the historical value. If the hash value changes, the symbol's content is determined to have been updated. A timestamp or version number-based detection method can also be used. For example, a timestamp or version number can be attached to the symbol's content, and the model compares the current version with the latest version upon each input. A deep detection method based on the semantics of the symbol's content can also be used. This involves performing semantic analysis of the symbol's content (such as natural language understanding or rule engine matching) and comparing the logical meanings of the old and new symbols to determine if they are consistent. If the logical meaning has changed, a change is determined.

[0051] Step 122 : Based on the perception content processing module of the shared layer, semantic features are extracted from the perception content to obtain perception semantic features.

[0052] Specifically, the perceptual content processing module can use a convolutional neural network (CNN) to extract features and a long short-term memory network (LSTM) to handle temporal dependencies. For example, the perceptual content processing module can first use a two-layer convolutional structure to process the input perceptual content, and then use a GRU (Gated Recurrent Unit) or LSTM network to handle the temporal dependencies of multiple perceptual contents. Furthermore, the two-layer convolutional structure in the perceptual content processing module can be replaced with optimized convolutional structures such as grouped convolution or point-wise convolution.

[0053] It's understandable that sensor data is generated by sensors. Since each sensor collects data at a different frequency, it's necessary to wait for each sensor to collect a data sequence before processing that data sequence. To address this, the sensory content processing module can be set to execute periodically at a set frequency. Specifically, the sensory content processing module extracts semantic features from the input sensory content at the set frequency. Furthermore, by adjusting the set frequency, it can achieve a balance between performance and energy consumption.

[0054] Specifically, when extracting semantic features from perceptual content through the perceptual content processing module, the first convolutional layer uses multiple small convolution kernels (e.g., 3×1 convolution) to extract local features from the input perceptual content. The output is then activated by a Reinforced Luminance (ReLU) activation function. The second convolutional layer then further integrates higher-order local features. The convolutional output is then expanded into a sequence along the time steps. The GRU unit controls the flow of information using update and reset gates, ultimately outputting the hidden state of the last time step or the mean of all time steps as temporal features. The local features output by the convolutional structure are fused with the temporal features of the GRU to produce perceptual semantic features.

[0055] Step 123 : Based on the fusion processing module of the shared layer, the symbolic semantic features and the perceptual semantic features are fused to obtain fused semantic features.

[0056] Specifically, the fusion processing module can use a fully connected layer to fuse the two features and semantics to obtain a fused semantic feature. For example, the fusion processing module can first merge the outputs of two preceding modules (i.e., the symbolic semantic features output by the symbolic content processing module and the perceptual semantic features output by the perceptual content processing module) by splicing, and then use a fully connected layer to process to obtain a fused semantic feature. Specifically, the symbolic semantic features and the perceptual semantic features are spliced ​​along the feature dimension, and then the spliced ​​features are mapped to a unified space through a fully connected layer (including an activation function such as ReLU) to generate a fused semantic feature for further processing by the task head.

[0057] Step 130 : Based on the task header of the energy management model, the fused semantic features are applied to generate a corresponding device control strategy, which is used to adjust the operating status of the electrical device.

[0058] Specifically, after receiving the fusion semantic features output by the fusion processing module, the task head can generate corresponding control decisions based on the features, that is, obtain the device control strategy. It should be noted that the specific implementation of step 130 can refer to the specific implementation of step 130 in the above embodiment and will not be repeated here.

[0059] Based on any of the above embodiments, Figure 3 This is the third flow chart of the energy management method provided by the present invention, such as Figure 3 As shown, the method includes: Step 110 , obtaining symbol content and perception content corresponding to the electrical device, wherein the symbol content includes external factors and / or user habits that affect the use of the electrical device, and the perception content includes device operation information and / or environmental perception information of the electrical device.

[0060] Specifically, the input of the energy management model includes two parts: symbolic content and perceptual content. The symbolic content is a converted symbol sequence, which includes two parts: one is the announcement information from external information sources, such as weather forecasts, extreme weather warnings, power supply status, grid load status, energy prices, etc. The announcement information is downloaded from the network in the form of a symbol sequence, and the conversion process is performed on the network server; the other is the locally recorded user habits, such as work and rest time, activity range, device association, etc. User habits are saved locally in the form of a symbol sequence.

[0061] Perception input is information acquired through sensors and electrical devices, such as sensor data, device operating status, family member behavior, and reservation requests and mode settings for large electrical devices. Perception input can be a fixed-size set of values, including data values ​​collected by numerical sensors, device operating status codes, family member behavior status codes collected by biometric sensors, and device reservation and mode status codes.

[0062] Step 120 , based on the shared layer of the energy management model, semantic features are extracted from the symbolic content and the perceptual content respectively, and the extracted symbolic semantic features and perceptual semantic features are fused to obtain fused semantic features corresponding to the electrical equipment.

[0063] It should be noted that the function of the shared layer is to extract common features and semantic information from the input. These extracted features and semantics are represented by data within the neural network. The shared layer consists of three submodules: the symbolic content processing module, the perceptual content processing module, and the fusion processing module. The symbolic content processing module specifically processes the symbolic content of the input, while the perceptual content processing module specifically processes the perceptual content of the input. These two modules initially extract features and semantics of the corresponding content, which are then input into the fusion processing module. The fusion processing module combines the preliminary features and semantics generated by these two modules and processes them into features and semantics that can be used by the task head.

[0064] Specifically, the above step 120 specifically includes: Step 121: extracting semantic features from the symbol content based on the symbol content processing module of the shared layer in the energy management model to obtain symbol semantic features; Step 122: extracting semantic features from the perceived content based on the perceived content processing module of the shared layer to obtain perceived semantic features; Step 123 : Based on the fusion processing module of the shared layer, the symbolic semantic features and the perceptual semantic features are fused to obtain fused semantic features.

[0065] Specifically, the symbol content processing module can use the attention mechanism to extract the semantic information of the symbol content, the perception content processing module can use a convolutional neural network (CNN) to extract features and a long short-term memory network (LSTM) to process temporal dependencies, and the fusion processing module can use a fully connected layer to fuse the two features and semantics. It should be noted that the specific implementation of steps 121, 122, and 123 can refer to the above embodiment and will not be repeated here.

[0066] Step 130 : Based on the task header of the energy management model, the fused semantic features are applied to generate a corresponding device control strategy, which is used to adjust the operating status of the electrical device.

[0067] It should be noted that the task head is the output layer for a specific task, and the number of task heads can be expanded. Each task head reads and further processes the features and semantics extracted by the shared layer to generate specific output for a specific task.

[0068] Specifically, the task head can use neural network methods to process features and semantics. Due to the different tasks of concern, different neural network methods can be adopted, such as deep neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory network (LSTM), attention mechanism, etc.

[0069] Specifically, different tasks can use different task heads, which are configured according to the specific task. For example, a task head can include a memory head and a control head. The memory head is a task head specifically used to record user habits, which can convert the identified user habits into a symbol sequence; the control head is a task head used to generate control strategies for electrical equipment. It can be divided into two categories: real-time tasks and planning tasks. The control heads can be divided according to electrical equipment and areas, and the electrical equipment controlled by different control heads can be non-overlapping. Here, the real-time task control head is used to predict the demand state (i.e., the expected state) of the electrical equipment and generate an immediate task event after comparing the actual state with the demand state; the planning task control head is used to generate a time-bound task event sequence (i.e., a planned task event sequence) based on the input appointment demand.

[0070] Furthermore, the task header includes a memory header and a control header. Accordingly, step 130 specifically includes: Step 131 , based on the memory head, according to a preset frequency period, the fusion semantic features are applied to identify the user habits, the identified user habits are converted into a user habit symbol sequence, and when the user habit symbol sequence changes, the pre-recorded user habit symbol sequence is updated.

[0071] Specifically, the memory head is a task head specifically designed to record user habits. It converts identified user habits into symbol sequences. Since user habits rarely change, there's no need to frequently update the user habit symbol sequence. In other words, the memory head can periodically generate detected user habits at a set frequency and dynamically update the locally stored user habit symbol sequence when changes occur, thereby reducing system energy consumption.

[0072] It is understandable that the memory head can be implemented using a Transformer decoder (Decoder), including six network layers: pre-order output multi-head self-attention, residual sum normalization, feature matrix multi-head self-attention, residual sum normalization, feedforward fully connected layer, residual sum normalization. Here, the pre-order output multi-head self-attention layer is used to perform self-attention calculations on the pre-order user habit symbol sequence to capture the dependency relationship of historical habits; the residual sum normalization layer is used to stabilize training; the feature matrix multi-head self-attention layer is used to align the historical user habit symbol sequence with the current fused semantic features; residual sum normalization, feedforward fully connected layer, and the final residual sum normalization layer are used to enhance feature expression capabilities. It should be understood that the memory head can also be implemented using the encoder structure of models such as BERT, ALBERT, and XLNet, and the embodiments of the present invention do not impose specific restrictions on this.

[0073] Specifically, the recorded user habit symbol sequence can be used as the pre-order output, and the fused semantic features output by the shared layer can be used as the feature matrix, which are input into the decoder together. The content output by the memory head is the updated user habit symbol sequence. Here, the input of the memory head includes two parts, one is the output of the shared layer, and the other is the feedback of its own output. It should be understood that the workflow of the memory head is: first, load the locally recorded user habit symbol sequence, then call the memory head at a preset frequency, input the user habit symbol sequence and the current fused semantic features into the decoder to generate a new user habit symbol sequence; compare the previously recorded user habit symbol sequence with the newly generated user habit symbol sequence. If there is a change, update the record and return the updated user habit symbol sequence.

[0074] Step 132 : Based on the control header, the fused semantic features are applied to generate corresponding device control strategies.

[0075] Specifically, a control head is a task head used to generate control strategies for electrical equipment. It is categorized into two types: real-time tasks and planned tasks. Real-time task heads generate immediate equipment operation events (also known as immediate task events), while planned task heads generate the expected task sequence and time nodes, i.e., the expected sequence of planned task events.

[0076] Furthermore, the control head includes a real-time task control head and a planned task control head, and the device control strategy includes an immediate task event and a planned task event sequence; accordingly, step 132 specifically includes: Step 1321 : Based on the real-time task class control header, the fusion semantic features are applied to predict the expected state of the electrical device, and an immediate task event is generated according to the expected state of the electrical device and the actual state of the electrical device.

[0077] Specifically, the structure of the real-time task control head can include three model layers: full connection, ReLU activation, and full connection. Among them, the first fully connected layer is used to map features to the hidden layer, and ReLU activation is used to introduce nonlinearity; the second fully connected layer is used to output the expected state value.

[0078] The workflow of a real-time task control head is as follows: the fused semantic features output by the shared layer are fed into a fully connected network to generate the expected state of the electrical device; the actual state of the device is then read (via sensors or device APIs); the actual state of the electrical device is compared with the expected state. When the actual state differs from the expected state, an immediate task event (such as "Start the air conditioner immediately") is generated. It should be understood that when training a real-time task control head, mean squared error can be used as the loss function to optimize the accuracy of state prediction.

[0079] It is understandable that in the home energy management scenario, areas can be divided according to the spatial characteristics of the actual application scenario, and a real-time task-type control head can be independently deployed in each area (such as living room, bedroom, kitchen, etc.) to avoid task conflicts and achieve area-level control.

[0080] Step 1322 : Based on the planned task class control header, the fusion semantic features are applied to generate an expected planned task event sequence. The planned task event sequence includes the execution order of each task event and the corresponding time nodes.

[0081] Specifically, the structure of the planning task control head can be implemented using a Transformer decoder. The existing unexecuted task event sequence is used as the pre-order output, and the features and semantics extracted by the shared layer are used as the feature matrix. These are then input into the decoder, resulting in the task event sequence after the newly added items are output by the planning task control head. It should be understood that the structure of the planning task control head can be consistent with the memory head, but the input is a task sequence rather than a sequence of user-familiar symbols.

[0082] The workflow of the task planning control head is as follows: first, load the existing unexecuted task event sequence (such as ["Start dishwasher at 18:00", "Start washing machine at 21:00"]); then, input the unexecuted task event sequence and the current fused semantic features into the decoder to generate the task event sequence after the newly added task (such as ["Start dishwasher at 18:00", "Start washing machine at 21:00", "Turn off air conditioner at 23:00"]).

[0083] Understandably, in a home energy management scenario, large household appliances can be clustered by function (e.g., cleaning equipment: dishwashers, dryers; entertainment equipment: TVs, stereos), and a planning task control head deployed for each category. For example, a cleaning task control head could be used to manage dishwashers, mobile cleaning robots, washing machines, dryers, etc.; an entertainment task control head could be used to manage TVs, stereos, etc.

[0084] In the embodiment of the present invention, the separation of user habit learning, real-time response and long-term planning is achieved through a hierarchical control head, so that the energy management model can take into account both flexibility and efficiency.

[0085] Based on any of the above embodiments, Figure 4 This is a schematic diagram of the structure of the energy management model provided by the present invention. Figure 4 As shown, the energy management model can be a distributed AI model that uses distributed multi-task learning to handle different control tasks. Figure 4 The distributed multi-task AI shown in the figure is the energy management model provided by an embodiment of the present invention. This model can be deployed on local distributed nodes. The model input mainly includes symbolic content and perceptual content, and the output is the control strategy for household appliances. The distributed AI model includes a shared layer for feature extraction and several dedicated task heads. The shared layer is executed collaboratively by all distributed nodes, while the task heads are executed individually by specific distributed nodes. The number of task heads can be expanded on demand.

[0086] The symbolic input consists of two parts: one is announcement information from external sources, such as weather forecasts, extreme weather warnings, power supply status, grid load status, and energy prices; the other is locally recorded user habits, such as sleep schedules, activity ranges, and device associations. The input perception content is information acquired through sensors and electrical devices, such as sensor data, device operating status, family member behavior, and reservation requests and mode settings for large electrical devices. The announcement information in the symbolic content is obtained through external communication and is public information. The user habits and perception content in the symbolic content are generated and acquired locally and are considered user privacy.

[0087] The output device control strategy is a series of tasks that change the operating state of electrical devices, such as "Set the bedroom air conditioner to 26°C for dehumidification" or "Schedule the washing machine to start at 11:00 PM." The control strategy here is an event instruction, not a control signal. The distributed AI model only generates the control strategy. The control signals for the specific electrical devices required to implement these event instructions are processed by other subsequent modules based on the specific device models and interfaces.

[0088] The shared layer extracts common features and semantic information from the input. These extracted features and semantics are represented as data within the neural network. The shared layer consists of three submodules: the symbolic content processing module, the perceptual content processing module, and the fusion processing module. The symbolic content processing module specifically processes the symbolic content of the input, while the perceptual content processing module specifically processes the perceptual content of the input. These two modules initially extract features and semantics of the corresponding content, which are then fed into the fusion processing module. The fusion processing module combines the preliminary features and semantics generated by these two modules, processing them into features and semantics that can be used by the task head. The symbolic content processing module uses an attention mechanism to extract semantic information from the symbolic content, while the perceptual content processing module uses a convolutional neural network (CNN) to extract features and a long short-term memory network (LSTM) to handle temporal dependencies. The fusion processing module uses a fully connected layer to fuse the two features and semantics.

[0089] The task head is the output layer for a specific task. The number of task heads is scalable. Each task head reads and further processes the features and semantics extracted by the shared layer to generate specific outputs for a specific task. Task heads use neural network methods to process features and semantics. Depending on the task, different neural network methods can be used. Possible neural network methods include deep neural networks (DNNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and attention mechanisms. Different tasks use different task heads, configured according to the specific task. For example, the memory head is specifically designed to record user habits, converting recognized user habits into symbol sequences. The control head is a task head responsible for making control decisions for electrical equipment. It is divided into real-time tasks and planning tasks. Tasks are divided according to electrical equipment and areas, and the electrical equipment controlled by different tasks do not overlap. The real-time control head predicts the demand state and generates task events after comparing the actual state with the demand state. The planning control head generates a timed sequence of task events based on the input reservation demand.

[0090] Based on any of the above embodiments, Figure 5 This is a schematic diagram of the overall process of the home energy management method based on the distributed AI model provided by the present invention. Figure 5 As shown in Figure 1, the process is divided into two stages: pre-deployment and post-deployment, which mainly include centralized pre-training, distributed inference, and fine-tuning. The pre-deployment stage mainly involves centralized pre-training.

[0091] Specifically, the energy management model is obtained through centralized pre-training before deployment and fine-tuning after deployment. The pre-training steps of the energy management model include: Based on the training data set, a multi-task learning method is used to pre-train each module of the shared layer in the energy management model to obtain the pre-trained shared layer; Based on the specific tasks of each task head in the energy management model, each task head is pre-trained, and a multi-task learning method is adopted to parallelly optimize the loss functions of different tasks on the pre-trained shared layer to synchronously adjust the pre-trained shared layer to obtain the pre-trained energy management model.

[0092] It should be noted that the purpose of centralized pre-training is to obtain a distributed AI model (i.e., an energy management model) that can be deployed. Centralized pre-training is performed on high-performance computing devices (such as workstations and cloud servers), and the shared layer neural network and the task head neural network are trained collaboratively.

[0093] Specifically, the shared layer neural network is first trained. Using a dedicated training dataset and a multi-task learning approach, the neural networks of each module in the shared layer are trained to enable them to extract common features and semantics from multi-channel and multi-type symbolic content and perceptual content. Next, the task head neural network is trained. Based on the specific tasks of each task head, the neural network of each task head is specially trained. With the help of multi-task learning methods, the loss functions of different tasks are optimized in parallel on the shared layer, and the shared layer model is adjusted synchronously. During the above training process, neural network optimization and compression techniques such as data quantization and model pruning can be used, and multiple rounds of iterative training can be performed to reduce the complexity and computational requirements of the model while maintaining the inference performance to the greatest extent.

[0094] It is understandable that each time the model is updated or the hardware system changes, the distributed AI model needs to be deployed to the distributed nodes. Deploying the model requires completing two tasks: one is to allocate computing tasks to each node based on the computing performance of each distributed node, and the other is to transmit the model and data control process to the corresponding distributed nodes. The model here includes the specific execution steps of the distributed AI model and data such as the weights required for each step. The data control process here includes the data interaction scheme between distributed nodes. The allocation method used in the embodiment of the present invention is static. After the allocation is completed, only one model transmission is required.

[0095] Furthermore, the post-deployment phase primarily involves distributed inference and fine-tuning. Distributed inference is designed to execute model inference operations on distributed hardware. Distributed inference involves three steps: collecting input data, extracting common features in the shared layer, and generating results in the task head. First, input data is collected by accessing network announcements, reading local user habits, recording sensor data, device operating status, and family member behavior, and converting them into the input data format required by the neural network. This data is then fed into the shared layer, where the symbolic content processing module and the perception content processing module process the corresponding content respectively. The fusion processing module generates common features and semantics. The symbolic content processing module executes only when the symbolic content changes, while the perception content processing module executes periodically at a set frequency. Next, the task head generates corresponding control decisions based on the features extracted by the shared layer. The memory head periodically generates detected user habits at a set frequency and updates the recorded user habit symbol sequence when changes occur. The real-time task control head generates immediate device operation events, and the planning task control head generates the expected task sequence and time nodes.

[0096] In addition, the fine-tuning of the energy management model is achieved based on online incremental learning after the model is deployed. The fine-tuning of the energy management model includes fine-tuning the network parameters of the fusion processing module and each control head in the energy management model, while the network parameters of the symbolic content processing module, perception content processing module and memory head in the energy management model are not fine-tuned.

[0097] Specifically, fine-tuning is intended to make the actually deployed model more intelligent and quickly adapt to the usage habits of specific users and the actual deployment environment. Fine-tuning primarily involves online incremental learning after deployment, performed on high-performance nodes in the distributed system. Based on information such as distributed AI decisions, sensor data feedback, and user behavior feedback, small adjustments are made to the neural network parameters of the fusion processing module and each control head in the model, allowing the distributed AI model to continuously learn the user's personalized characteristics. The symbolic content processing module, the perception content processing module, and the neural network of the memory head do not participate in fine-tuning. According to the set batch processing capacity, the weights are updated once after sufficient data is collected. According to the set frequency, the weights are updated several times and then redeployed to each node. This is because when fine-tuning slightly, a batch of fine-tuning data needs to be accumulated to better reflect it on the model. If the weights are updated once and then deployed, almost no effect will be seen.

[0098] The method proposed in the embodiment of the present invention can make automated decisions on various energy management and household appliance control tasks, helping to reduce energy costs; it can autonomously learn the relationship between user habits and household appliances, with a high degree of personalization and strong adaptability; it is highly flexible by utilizing multi-channel and multi-type information to assist work; it executes computing tasks locally without relying on external data transmission, with strong reliability; and it does not require the external exchange of private data, effectively protecting privacy and security.

[0099] The energy management system provided by the present invention is described below. The energy management system described below and the energy management method described above can be referenced to each other.

[0100] Based on any of the above embodiments, Figure 6 This is a schematic diagram of the structure of the energy management system provided by the present invention. Figure 6 As shown, the system includes: Content acquisition module 610, for acquiring symbolic content and perceptual content corresponding to an electrical device, wherein the symbolic content includes external factors affecting the use of the electrical device and / or user habits, and the perceptual content includes device operation information and / or environmental perception information of the electrical device; Feature extraction module 620 is used to extract semantic features from the symbolic content and the perceptual content based on the shared layer of the energy management model, and fuse the extracted symbolic semantic features and perceptual semantic features to obtain a fused semantic feature corresponding to the electrical device; The strategy generation module 630 is used to generate corresponding device control strategies based on the task header of the energy management model and the fusion semantic features. The device control strategies are used to adjust the operating status of electrical devices.

[0101] The system provided by an embodiment of the present invention obtains the symbolic content and sensory content corresponding to electrical devices. The symbolic content integrates external factors and / or user habits that influence appliance usage, providing a macro-level decision-making basis for energy management. The sensory content includes real-time collected device operating information and environmental perception information, ensuring that decisions are close to actual scenarios. Semantic features are extracted from the symbolic and sensory content separately through the shared layer of the energy management model, capturing the deep semantic connections between different information sources and improving the ability to understand complex scenarios. Semantic feature extraction and fusion transform discrete multi-source information into structured features, not only avoiding the limitations of a single information source and achieving global optimization of energy management, but also organically integrating symbolic logic with sensory data, enabling decisions to take into account user habits and real-time needs, reducing misjudgments and improving the accuracy, adaptability, and personalization of energy management. Furthermore, by generating device control strategies based on fused semantic features, the model can respond to changes in external factors or environmental conditions in real time, dynamically optimizing control strategies and avoiding static, "one-size-fits-all" rules, thereby improving energy efficiency.

[0102] Based on any of the above embodiments, the feature extraction module 620 includes: A symbol content processing unit, configured to extract semantic features from the symbol content based on the symbol content processing module of the shared layer to obtain the symbol semantic features; a perception content processing unit, configured to extract semantic features from the perception content based on the perception content processing module of the shared layer to obtain the perception semantic features; A fusion processing unit is used to fuse the symbolic semantic feature and the perceptual semantic feature based on the fusion processing module of the shared layer to obtain the fused semantic feature.

[0103] Based on any of the above embodiments, the symbol content processing unit is specifically configured to: Based on the symbol content processing module, when a change in the symbol content is detected, the attention mechanism is used to extract semantic features of the symbol content to obtain the symbol semantic features.

[0104] Based on any of the above embodiments, the task header includes a memory header and a control header, and the strategy generation module 630 includes: a memory head unit, configured to identify user habits based on the memory head and using the fused semantic features at a preset frequency period, convert the identified user habits into a user habit symbol sequence, and update the pre-recorded user habit symbol sequence when the user habit symbol sequence changes; A control head unit is used to apply the fused semantic features based on the control head to generate a corresponding device control strategy.

[0105] Based on any of the above embodiments, the control head includes a real-time task control head and a planned task control head, and the device control strategy includes an immediate task event and a planned task event sequence; accordingly, the control head unit is specifically used to: Based on the real-time task class control header, applying the fused semantic features, predicting the expected state of the electrical device, and generating an immediate task event according to the expected state of the electrical device and the actual state of the electrical device; Based on the planned task class control header, the fused semantic features are applied to generate an expected planned task event sequence, where the planned task event sequence includes the execution order of each task event and the corresponding time nodes.

[0106] Based on any of the above embodiments, the energy management model is obtained based on centralized pre-training before deployment and fine-tuning after deployment. The system further includes a pre-training unit, which is configured to: Based on the training data set, a multi-task learning method is used to pre-train each module of the shared layer in the energy management model to obtain a pre-trained shared layer; Based on the specific tasks of each task head in the energy management model, each task head is pre-trained, and a multi-task learning method is adopted to parallelly optimize the loss functions of different tasks on the pre-trained shared layer to synchronously adjust the pre-trained shared layer to obtain a pre-trained energy management model.

[0107] Based on any of the above embodiments, the fine-tuning of the energy management model is achieved based on online incremental learning after the model is deployed. The fine-tuning of the energy management model includes fine-tuning the network parameters of the fusion processing module and each control head in the energy management model, and does not fine-tune the network parameters of the symbolic content processing module, the perception content processing module and the memory head in the energy management model.

[0108] Based on any of the above embodiments, the energy management model is deployed on a local distributed node, and the user habits in the symbolic content and the perception content are generated and acquired locally.

[0109] Figure 7 An example of a physical structure diagram of an electronic device is shown below. Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other via the communication bus 740. The processor 710 may call logic instructions in the memory 730 to execute an energy management method, which includes: obtaining symbolic content and perceptual content corresponding to an electrical device, wherein the symbolic content includes external factors and / or user habits that affect the use of the electrical device, and the perceptual content includes device operation information and / or environmental perception information of the electrical device; based on the shared layer of the energy management model, performing semantic feature extraction on the symbolic content and the perceptual content respectively, and fusing the extracted symbolic semantic features and perceptual semantic features to obtain a fused semantic feature corresponding to the electrical device; based on the task head of the energy management model, applying the fused semantic features to generate a corresponding device control strategy, wherein the device control strategy is used to adjust the operating state of the electrical device.

[0110] Furthermore, the logic instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the relevant art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0111] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the energy management method provided by the above methods, which includes: obtaining the symbolic content and perception content corresponding to the electrical equipment, the symbolic content includes external factors and / or user habits that affect the use of the electrical equipment, and the perception content includes the equipment operation information and / or environmental perception information of the electrical equipment; based on the shared layer of the energy management model, semantic features are extracted for the symbolic content and perception content respectively, and the extracted symbolic semantic features and perception semantic features are fused to obtain the fused semantic features corresponding to the electrical equipment; based on the task head of the energy management model, the fused semantic features are applied to generate a corresponding equipment control strategy, and the equipment control strategy is used to adjust the operating status of the electrical equipment.

[0112] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the energy management method provided by the above-mentioned methods, the method comprising: obtaining the symbolic content and perception content corresponding to the electrical equipment, the symbolic content including external factors and / or user habits that affect the use of the electrical equipment, and the perception content including the equipment operation information and / or environmental perception information of the electrical equipment; based on the shared layer of the energy management model, performing semantic feature extraction on the symbolic content and the perception content respectively, and fusing the extracted symbolic semantic features and perception semantic features to obtain the fused semantic features corresponding to the electrical equipment; based on the task head of the energy management model, applying the fused semantic features to generate a corresponding equipment control strategy, and the equipment control strategy is used to adjust the operating status of the electrical equipment.

[0113] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0114] Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the relevant technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An energy management method, characterized in that: include: Acquiring symbolic content and perceptual content corresponding to an electrical device, wherein the symbolic content includes external factors and / or user habits that affect the use of the electrical device, and the perceptual content includes device operation information and / or environmental perception information of the electrical device; Based on the shared layer of the energy management model, semantic features of the symbol content and the perception content are extracted respectively, and the extracted symbol semantic features and perception semantic features are fused to obtain fused semantic features corresponding to the electrical device; Based on the task head of the energy management model, the fused semantic features are applied to generate a corresponding device control strategy, where the device control strategy is used to adjust the operating state of the electrical device.

2. The energy management method according to claim 1, characterized in that: The shared layer based on the energy management model extracts semantic features from the symbol content and the perception content respectively, and fuses the extracted symbol semantic features and perception semantic features to obtain a fused semantic feature corresponding to the electrical device, including: Based on the symbol content processing module of the shared layer, semantic feature extraction is performed on the symbol content to obtain the symbol semantic feature; Extracting semantic features of the perceived content based on the perceived content processing module of the shared layer to obtain the perceived semantic features; Based on the fusion processing module of the shared layer, the symbolic semantic feature and the perceptual semantic feature are fused to obtain the fused semantic feature.

3. The energy management method according to claim 2, characterized in that: The symbol content processing module based on the shared layer extracts semantic features from the symbol content to obtain the symbol semantic features, including: Based on the symbol content processing module, when a change in the symbol content is detected, the attention mechanism is used to extract semantic features of the symbol content to obtain the symbol semantic features.

4. The energy management method according to claim 1, characterized in that: The task head includes a memory head and a control head. The task head based on the energy management model applies the fused semantic features to generate a corresponding device control strategy, including: Based on the memory head, applying the fused semantic features to identify user habits according to a preset frequency period, converting the identified user habits into a user habit symbol sequence, and updating the pre-recorded user habit symbol sequence when the user habit symbol sequence changes; Based on the control head, the fused semantic features are applied to generate corresponding device control strategies.

5. The energy management method according to claim 4, characterized in that: The control head includes a real-time task control head and a planned task control head, and the device control strategy includes an immediate task event and a planned task event sequence; Accordingly, the generating of a corresponding device control strategy based on the control head and applying the fused semantic features includes: Based on the real-time task class control header, applying the fused semantic features, predicting the expected state of the electrical device, and generating an immediate task event according to the expected state of the electrical device and the actual state of the electrical device; Based on the planned task class control header, the fused semantic features are applied to generate an expected planned task event sequence, where the planned task event sequence includes the execution order of each task event and the corresponding time nodes.

6. The energy management method according to any one of claims 1 to 5, characterized in that: The energy management model is obtained based on centralized pre-training before deployment and fine-tuning after deployment. The pre-training steps of the energy management model include: Based on the training data set, a multi-task learning method is used to pre-train each module of the shared layer in the energy management model to obtain a pre-trained shared layer; Based on the specific tasks of each task head in the energy management model, each task head is pre-trained, and a multi-task learning method is adopted to parallelly optimize the loss functions of different tasks on the pre-trained shared layer to synchronously adjust the pre-trained shared layer to obtain a pre-trained energy management model.

7. The energy management method according to claim 6, characterized in that: The fine-tuning of the energy management model is achieved based on online incremental learning after the model is deployed. The fine-tuning of the energy management model includes fine-tuning the network parameters of the fusion processing module and each control head in the energy management model, and does not fine-tune the network parameters of the symbolic content processing module, the perception content processing module and the memory head in the energy management model.

8. The energy management method according to any one of claims 1 to 5, characterized in that: The energy management model is deployed on a local distributed node, and the user habits in the symbolic content and the perception content are generated and acquired locally.

9. An energy management system, characterized in that: include: a content acquisition module, configured to acquire symbolic content and perceptual content corresponding to an electrical device, wherein the symbolic content includes external factors and / or user habits that affect the use of the electrical device, and the perceptual content includes device operation information and / or environmental perception information of the electrical device; a feature extraction module for extracting semantic features of the symbolic content and the perceptual content based on the shared layer of the energy management model, and fusing the extracted symbolic semantic features and perceptual semantic features to obtain a fused semantic feature corresponding to the electrical device; A strategy generation module is used to generate a corresponding device control strategy based on the task header of the energy management model and the application of the fusion semantic features, wherein the device control strategy is used to adjust the operating state of the electrical device.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the energy management method according to any one of claims 1 to 8 is implemented.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the energy management method according to any one of claims 1 to 8 is implemented.

12. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the energy management method according to any one of claims 1 to 8 is implemented.

Images