A big data processing system for socket energy efficiency optimization and energy consumption prediction

Abstract

This invention relates to the field of data processing technology and discloses a big data processing system for socket energy efficiency optimization and energy consumption prediction. The system includes at least one socket and a cloud server. The socket's edge computing unit converts real-time collected device electrical signals into behavioral morpheme sequences through an appliance morpheme extraction module. The cloud server receives these sequences, and its sequence prediction module predicts future electricity consumption behavior, energy consumption, and dependencies. A decision state generation module integrates the predicted information to construct a comprehensive state vector. An active intervention decision module generates and issues optimal energy efficiency intervention commands based on this vector. This invention achieves accurate energy consumption prediction and energy efficiency optimization for socket electricity consumption. Through deep analysis and attribution understanding of electricity consumption behavior, combined with multi-objective optimization for intelligent decision-making, it can generate a comprehensively optimal intervention strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a socket energy efficiency optimization and energy consumption prediction big data processing system. BACKGROUND

[0002] With the development of smart grid and Internet of Things technology, the demand for fine energy management in society is growing. As a key node connecting the end of the power grid and electrical equipment, smart sockets are an important carrier for realizing energy consumption data collection and electrical behavior control in home and office scenarios. By monitoring and analyzing electrical signals at the socket level, a technical premise is provided for energy efficiency optimization of individual electrical equipment.

[0003] Currently, energy-saving applications around smart sockets usually focus on providing real-time power metering and basic remote or timed control. Some solutions will make simple threshold judgments based on collected power data, such as automatically cutting off power when detecting that the electrical equipment enters standby state. These applications have improved the convenience of electricity management, but their technical core still remains at the shallow use of electrical signal data.

[0004] However, the existing technology simplifies the data processing method to statistics or comparison of power values, losing the rich time sequence and frequency domain features in the original electrical signals that can represent the specific running state of the equipment, so it cannot build a deep understanding of the fine behavior of the equipment, and it cannot effectively predict future electricity consumption behavior and energy consumption trends. Therefore, its energy efficiency management strategy can only be a lagging, passive response control. Ultimately, due to the lack of prediction ability and understanding of the context of complex electricity consumption behavior, its optimization strategy often relies on static rules, making it difficult to balance between dynamic electricity prices, user needs, and equipment working conditions, resulting in limited overall optimization effect.

[0005] Therefore, the present application proposes a socket energy efficiency optimization and energy consumption prediction big data processing system to solve the deficiencies of the prior art. SUMMARY

[0006] The prior art has deficiencies in energy efficiency optimization and energy consumption prediction of sockets. First, at the big data processing level, it cannot extract deep patterns representing fine behavior of equipment from massive electrical signals. Second, due to the lack of data processing capability, it cannot accurately predict future trends at the energy consumption prediction level, resulting in a management strategy that can only respond passively. Finally, at the energy efficiency optimization level, its strategy is based on static rules and cannot combine prediction results and electricity context, making it difficult to balance between electricity prices, user preferences, and equipment working conditions, resulting in poor optimization effect and adaptability.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The first aspect of the application provides a large data processing system for socket energy efficiency optimization and energy consumption prediction, comprising: at least one socket and a cloud server, wherein the socket is provided with an edge computing unit, and the cloud server is in communication connection with the edge computing unit.

[0009] The edge computing unit comprises:

[0010] An appliance morpheme extraction module is configured to collect raw electrical signals of an electrical equipment connected to the socket in real time, and process the raw electrical signals into a sequence of appliance morphemes arranged in time sequence and representing the basic electrical behavior of the electrical equipment.

[0011] The cloud server comprises:

[0012] A sequence prediction module is configured to receive the sequence of appliance morphemes, and based on the historical context of the sequence of appliance morphemes, predict the electrical behavior in a future period of time, and generate a corresponding future sequence of appliance morphemes, an energy consumption trend prediction result, and an attention weight representing the dependency relationship between historical appliance morphemes.

[0013] A decision state generation module is configured to generate a comprehensive state vector comprising electrical equipment behavior context information by combining the energy consumption trend prediction result and the attention weight.

[0014] An active intervention decision module is configured to receive the comprehensive state vector, and output an energy efficiency optimization intervention instruction for the socket based on the comprehensive state vector.

[0015] As a preferred technical solution, the appliance morpheme extraction module is specifically operated in the following manner: the raw electrical signals are collected at a preset frequency to form a signal window; the signal window is subjected to time-frequency domain transformation to generate a time-frequency spectrum; a preset neural network serving as a convolutional neural network is used to analyze and classify the time-frequency spectrum to map each signal window to an appliance morpheme. This process can be represented by the following formula:

[0016] m t =CNN(S t );

[0017] In the formula, S t is the time-frequency spectrum generated at time t; CNN(·) represents a classification function performed by the convolutional neural network; m t is the appliance morpheme output at the corresponding time. Thus, a sequence of appliance morphemes arranged in time sequence {m1,m2,...,m t} is generated.

[0018] As a preferred technical solution, the sequence prediction module is operated by an encoder-decoder model based on a Transformer architecture. The encoder part of the model processes the historical context of the appliance token sequence, and the decoder part of the model generates the future appliance token sequence. The historical context of the appliance token sequence is a preset length of appliance token sub-sequence before the time point of making a prediction. The model further combines the future appliance token sequence with a preset mapping relationship between appliance tokens and power consumption to generate the energy consumption trend prediction result. The attention weight is calculated by a multi-head self-attention mechanism built in the encoder-decoder model.

[0019] As a preferred technical solution, the decision state generation module is specifically operated in the following manner: based on the attention weight, the historical appliance tokens in the appliance token sequence are subjected to attribution analysis to identify key historical appliance tokens that have a decisive effect on the current state; according to the result of the attribution analysis, a context criticality vector quantifying the influence degree of the key historical appliance tokens is generated; and the context criticality vector and the energy consumption trend prediction result are spliced to generate the comprehensive state vector including appliance behavior context information.

[0020] As a further preferred technical solution, the decision state generation module, when generating the comprehensive state vector, is further configured to splice an external environment information vector and a user preference vector, the external environment information vector including at least one of real-time electricity price information and power grid load state, and the user preference vector including at least one of historical intervention instruction acceptance rate and user preset mode preference. The splicing process can be represented by the following formula:

[0021] S t =concat(V ctx ,R pred ,V env ,V user );

[0022] In the formula, S t is the comprehensive state vector generated at time t; concat(·) represents a vector splicing operation; V ctx is the context criticality vector; R pred is the energy consumption trend prediction result; V env is the external environment information vector; and V user is the user preference vector.

[0023] As a preferred technical solution, the active intervention decision module runs by using a reinforcement learning agent which is a pre-trained deep Q network model, inputs the comprehensive state vector as a current state, and selects an action which can maximize long-term cumulative reward by the reinforcement learning agent to generate the energy efficiency optimization intervention instruction. The selection process of the action can be represented by the following formula:

[0024] a t =argmax a Q(S t ,a;θ);

[0025] In the formula, a t is the action selected at time t, i.e. the generated energy efficiency optimization intervention instruction; a is an action in the set of all possible actions; Q(S t ,a;θ) is the value function of the deep Q network performing the action a under the state S t ; and θ is the parameter of the network.

[0026] As a further preferred technical solution, the decision basis of the active intervention decision module outputting the energy efficiency optimization intervention instruction is a multi-objective reward function, which at least includes evaluation of three dimensions of expected energy saving amount, user satisfaction and device health degree. The reward function can be represented by the following formula:

[0027] R t =w e ·E t +w u ·U t +w d ·D t ;

[0028] In the formula, R t is the immediate reward at time t; E t , U t , and D t represent the quantitative values of expected energy saving amount, user satisfaction and device health degree respectively; w e , w u , and w d are corresponding weight coefficients.

[0029] As a preferred technical solution, the cloud server further comprises a dynamic morpheme library updating module, which is configured to generate a new definition of the power consumption device morpheme when the sequence prediction module identifies a new power consumption behavior pattern which cannot be accurately described by the existing power consumption device morpheme, and to issue the new definition of the power consumption device morpheme to the appliance morpheme extraction module to dynamically expand the power consumption device morpheme library.

[0030] The second aspect of the application provides a big data processing method for socket energy efficiency optimization and energy consumption prediction, which is applied to the system and comprises the following steps:

[0031] Real-time acquisition of original electric signals of the electric equipment connected to the socket, and processing of the original electric signals into a time-sequenced electric equipment morpheme sequence representing the basic electric behavior of the electric equipment;

[0032] Receiving the electric equipment morpheme sequence, and predicting the electric behavior in a future period of time based on the historical context of the electric equipment morpheme sequence, and generating corresponding future electric equipment morpheme sequence, energy consumption trend prediction result, and attention weight representing the dependency relationship between historical electric equipment morphemes;

[0033] Combining the energy consumption trend prediction result and the attention weight, generating a comprehensive state vector including electric equipment behavior context information;

[0034] Receiving the comprehensive state vector, and outputting an energy efficiency optimization intervention instruction for the socket based on the comprehensive state vector.

[0035] The application provides a big data processing system for socket energy efficiency optimization and energy consumption prediction, which has the following beneficial effects:

[0036] 1. The application processes the original electric signal big data stream collected by the socket through the electric appliance morpheme extraction module and the sequence prediction module, converts it into a structured electric equipment morpheme sequence, and performs time sequence prediction. This enables the system to accurately predict future energy consumption trends, realizes the transition from passive response to active prediction, and provides a key decision basis for subsequent energy efficiency optimization scheduling.

[0037] 2. The application uses the attention weight to perform attribution analysis on the historical electric equipment morphemes through the decision state generation module, thereby generating a comprehensive state vector containing accurate context information. This enables the system not only to predict the future, but also to understand the key historical behavior leading to the prediction, significantly improving the pertinence and effectiveness of the subsequent energy efficiency optimization intervention instruction.

[0038] 3. The application adopts an active intervention decision module based on reinforcement learning, and configures a multi-objective reward function including expected energy consumption saving amount, user satisfaction and equipment health degree. This enables the system to automatically weigh between multiple dimensions when making energy efficiency optimization decisions, and find the most comprehensive intervention strategy, surpassing the traditional simple rule system with only energy saving as a single target. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 It is a big data processing system architecture diagram for socket energy efficiency optimization and energy consumption prediction according to the application.

[0040] Figure 2 The electric appliance morpheme extraction flowchart according to the present application;

[0041] Figure 3 The internal structure diagram of the sequence prediction module according to the present application;

[0042] Figure 4 The internal processing flowchart of the decision state generation module according to the present application;

[0043] Figure 5 The operation flowchart of the active intervention decision module according to the present application;

[0044] Figure 6 The operation flowchart of the dynamic morpheme library updating module according to the present application;

[0045] Figure 7 The big data processing method flowchart of the socket energy efficiency optimization and energy consumption prediction according to the present application.

[0046] Wherein, 10, edge computing unit; 110, electric appliance morpheme extraction module; 20, cloud server; 210, sequence prediction module; 220, decision state generation module; 230, active intervention decision module; 240, dynamic morpheme library updating module. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the present application specification. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0048] Referring to the drawings Figure 1 , Figure 1 The socket energy efficiency optimization and energy consumption prediction big data processing system architecture diagram according to one embodiment of the present application. The present application provides a socket energy efficiency optimization and energy consumption prediction big data processing system, which comprises at least one socket (not shown in the figure) and a cloud server 20. An edge computing unit 10 is arranged in the socket, and the edge computing unit 10 and the cloud server 20 are connected through communication modes such as Wi-Fi, Zigbee or mobile communication network.

[0049] The edge computing unit 10 comprises an electric appliance morpheme extraction module 110.

[0050] The cloud server 20 comprises a sequence prediction module 210, a decision state generation module 220, an active intervention decision module 230, and a dynamic n-gram library updating module 240.

[0051] Referring to the drawings Figure 2 , Figure 2 is a flowchart of the appliance n-gram extraction according to an embodiment of the present application. The flow is executed by the appliance n-gram extraction module 110 in the edge computing unit 10 installed in the socket.

[0052] In a specific embodiment, the appliance n-gram extraction module 110 is configured to convert the continuous, unstructured raw electrical signals generated by the electrical device into a discrete, structured time series data, i.e. the appliance n-gram sequence, to provide high-quality input for the deep analysis and prediction of the cloud server 20.

[0053] The module first acquires the raw electrical signals of the electrical device flowing through the socket in real time by the built-in acquisition circuit at a preset sampling frequency (e.g. 10 kHz or higher), which can be the current signal, the voltage signal, or a combination of the two. The high sampling frequency ensures that the transient characteristics representing the device state switching can be captured.

[0054] The acquired continuous signal stream is then segmented into signal windows with fixed time length (e.g. 2 seconds) arranged in time series. For each signal window, the module performs time-frequency domain transformation processing to generate a time-frequency spectrogram. In a specific implementation, the time-frequency domain transformation method used is the short-time Fourier transform (STFT). The generated time-frequency spectrogram is a two-dimensional matrix that can show how the frequency components of the signal change over time within the signal window time period.

[0055] Subsequently, the appliance n-gram extraction module 110 uses a pre-set neural network model as a convolutional neural network (CNN) to analyze and classify the time-frequency spectrogram generated in the previous step. Since the time-frequency spectrogram is structurally similar to image data, the use of a convolutional neural network can effectively extract its local and global texture features, which have high correlation with the specific operating state of the electrical device (such as start, stop, constant power, variable frequency, etc.).

[0056] The convolutional neural network model receives a time-frequency spectrogram as input and outputs a classification result, which is defined as an electrical device n-gram. This process can be represented by the following formula:

[0057] m t =CNN(S t );

[0058] In the formula, S tis the time-frequency spectrogram generated at time t; CNN(·) represents a classification function performed by a convolutional neural network; m t is the outputted appliance morpheme at the corresponding time. For example, the morpheme can be one of “compressor start”, “standby state”, “heating module constant power running”, or “motor idling”, etc.

[0059] By sequentially performing the above-mentioned acquisition, transformation and classification processes on consecutive signal windows, the appliance morpheme extraction module 110 finally generates an appliance morpheme sequence {m1, m2,..., m t} arranged in time sequence. The sequence is then sent to the cloud server 20 for subsequent processing through the communication link between the edge computing unit 10 and the cloud server 20.

[0060] Referring to the accompanying Figure 3 , Figure 3 is a schematic diagram of the internal structure of a sequence prediction module according to an embodiment of the present application. The sequence prediction module 210 provided in the cloud server 20 is used to receive the appliance morpheme sequence generated by the appliance morpheme extraction module 110 and perform a sequence-to-sequence prediction task.

[0061] In a specific embodiment, the core of the sequence prediction module 210 is an encoder-decoder model based on the Transformer architecture. The module first converts the received discrete appliance morpheme sequence into a sequence of continuous real vectors through a word embedding layer (Embedding#Layer) to facilitate processing by the neural network model.

[0062] The encoder part of the model is responsible for processing the historical context of the appliance morpheme sequence. The historical context is defined as a sub-sequence of appliance morphemes of a preset length before the time point at which the prediction is made. The encoder encodes the input historical context vector sequence through its internal multi-layer stacking structure to generate a set of intermediate representation vectors that can represent the deep semantic information of the historical context.

[0063] Inside the encoder, the core operation unit is the multi-head self-attention mechanism (Multi-Head#Self-Attention). This mechanism enables the encoder to calculate the degree of association between any one position in the sequence and all other morphemes in the sequence when processing the morpheme at that position, and this degree of association is quantified as an attention weight. The calculation process of the attention weight can be represented by the following formula:

[0064]

[0065] where Q, K, V represent Query, Key and Value matrix respectively, which are linearly transformed from the input vector sequence. k is the dimension of the key vector. The result of this formula is the value vector after weighted summation, which reflects the importance of different parts of the sequence to the current part. The attention score matrix generated in this calculation process is the attention weight representing the dependency between historical appliance morphemes, which is output to the decision state generation module 220.

[0066] The decoder part of the model receives the intermediate representation vector output by the encoder and generates appliance morphemes at future time steps one by one in an autoregressive manner. At each time step, the decoder combines the output of the encoder and the morpheme generated at the previous time step to predict the most likely morpheme at the current time step, ultimately forming a complete sequence of future appliance morphemes.

[0067] After generating the sequence of future appliance morphemes, the sequence prediction module 210 further combines the future sequence with a pre-set mapping relationship between appliance morphemes and power consumption to generate the energy consumption trend prediction result. The mapping relationship can be a lookup table that stores the average power consumption value corresponding to each morpheme, or a small regression model. By converting each morpheme in the future sequence into the corresponding power consumption value, the module can output a detailed energy consumption curve for a future period of time, i.e., the energy consumption trend prediction result.

[0068] Referring to the accompanying Figure 4 , Figure 4 is a schematic diagram of the internal processing flow of the decision state generation module according to an embodiment of the present application. The decision state generation module 220 set in the cloud server 20 is used to integrate the prediction information from the sequence prediction module 210 and the context information, and combine external data to construct a comprehensive information state vector for subsequent decision-making.

[0069] In a specific embodiment, the decision state generation module 220 first receives two core data output from the sequence prediction module 210: the energy consumption trend prediction result and the attention weight.

[0070] Based on the attention weight, the module performs attribution analysis on the historical appliance morphemes in the sequence of appliance morphemes. Since the attention weight matrix quantifies the contribution of each morpheme in the historical sequence to the generation of the prediction result, by analyzing the matrix, one or more key historical appliance morphemes that have a decisive effect on the current and future state can be identified.

[0071] According to the result of the attribution analysis, the module generates a context criticality vector. The vector is a quantitative description of the key historical appliance n-gram and its impact degree. For example, if the analysis finds that the historical n-gram “compressor start” has the highest contribution to the high energy consumption prediction result in the future period, the context criticality vector will have a higher value in the corresponding dimension, thereby explicitly representing the key causal relationship in the state vector.

[0072] After the context criticality vector is generated, the decision state generation module 220 performs a vector splicing operation to generate the final comprehensive state vector. In one specific implementation, the module splices the context criticality vector with the energy consumption trend prediction result.

[0073] Further, in order to make the decision more comprehensive, the module can also obtain an external environment information vector and a user preference vector from an external data interface, and splice them together. The external environment information vector includes at least one of real-time electricity price information and grid load state. The user preference vector includes at least one of the intervention instruction acceptance rate statistically obtained from historical interactions and the user's preset mode preference (such as energy saving mode or performance mode).

[0074] The complete splicing process can be characterized by the following formula:

[0075] S t =concat(V ctx ,R pred ,V env ,V user );

[0076] In the formula, S t is the comprehensive state vector generated at time t, which will be the input of the next module, i.e., the active intervention decision module 230; concat(·) represents the vector splicing operation; V ctx is the context criticality vector; R pred is the energy consumption trend prediction result; V env is the external environment information vector; and V user is the user preference vector.

[0077] Referring to the accompanying drawings, Figure 5 , Figure 5 is a schematic diagram of the operation flow of the active intervention decision module according to one embodiment of the present application. The active intervention decision module 230 provided in the cloud server 20 is configured to receive the comprehensive state vector generated by the decision state generation module 220, and output a specific energy efficiency optimization intervention instruction for the socket based on the vector.

[0078] In one specific embodiment, the module is implemented by a reinforcement learning agent in the form of a pre-trained deep Q-network (DQN) model. The reinforcement learning agent takes the composite state vector S t as the current state input.

[0079] Based on this state input, the agent evaluates the long-term cumulative reward that can be achieved by performing each possible action in the current state using its internal value function network, and selects the action that maximizes the expected reward to generate the energy efficiency optimization intervention instruction. This action selection process can be characterized by the following equation:

[0080] a t = argmax a Q(S t ,a; θ);

[0081] where a t is the action selected at time t, i.e., the generated energy efficiency optimization intervention instruction, such as "delayed start", "low-power operation", or "immediate shutdown", etc.; a is an action in the set of all possible actions; Q(S t ,a; θ) is the value function of the deep Q-network for performing action a in state S t ; and θ is the parameter of the network.

[0082] The basis for the decision of the reinforcement learning agent is the multi-objective reward function used in its pre-training process. This reward function is used to provide quantitative feedback signals for the immediate results produced by the agent after performing an action in a specific state during the training phase. In one specific implementation, the reward function includes at least the evaluation of the expected energy saving amount, user satisfaction, and device health. This reward function can be characterized by the following equation:

[0083] R t = w e · E t + w u · U t + w d · D t ;

[0084] where R t is the immediate reward at time t; E t , U t , and D t represent the quantitative values of the expected energy saving amount, user satisfaction, and device health, respectively; w e , w u , and w d are the corresponding weight coefficients used to adjust the relative importance between different optimization objectives, so that the final decision can achieve a balance between multiple dimensions.

[0085] Referring to the drawings Figure 6 , Figure 6 is a schematic diagram of the operation flow of the dynamic morpheme library updating module according to an embodiment of the present application. The dynamic morpheme library updating module 240 in the cloud server 20 is used to realize adaptive expansion of the morpheme library of the electrical equipment, so as to ensure that the system can recognize and process the newly emerging electrical behavior mode.

[0086] In a specific embodiment, the updating flow of the module is triggered by a specific condition. When the sequence prediction module 210 is processing a morpheme sequence of an electrical equipment, if the prediction output of the future morpheme has low confidence on all known morpheme categories, or when the edge computing unit 110 reports that the time-frequency spectrogram cannot be effectively classified by the convolutional neural network, the system determines that a new electrical behavior mode that cannot be accurately described by the existing electrical equipment morpheme has occurred, and at this time, the operation of the dynamic morpheme library updating module 240 is triggered.

[0087] After being triggered, the module first retrieves the original signal window data and its time-frequency spectrogram corresponding to the new electrical behavior mode from the data record. The module can accumulate multiple time-frequency spectrogram samples marked as unknown. Subsequently, the module performs an unsupervised clustering algorithm (for example, DBSCAN or K-Means algorithm) on the accumulated set of unknown time-frequency spectrogram samples, and aggregates samples with similar features together. Each generated stable cluster is considered as a candidate new electrical equipment morpheme.

[0088] For each candidate new electrical equipment morpheme, the module generates a new morpheme definition for it. The definition includes assigning a unique new morpheme identifier and extracting the feature representation of the cluster (for example, the centroid of all time-frequency spectrograms in the cluster or a representative sample).

[0089] Then, the module uses these sample data with new morpheme identifiers to update or fine-tune a baseline convolutional neural network model deployed on the cloud server 20. This process integrates the new electrical behavior mode and its corresponding morpheme definition into the classification model.

[0090] After the model updating and verification are completed in the cloud, the dynamic morpheme library updating module 240 distributes the updated convolutional neural network model parameters to one or more specified edge computing units 10 through the cloud-edge communication link. After the electrical appliance morpheme extraction module 110 in the edge computing unit 10 receives the new model parameters, it replaces the local preset neural network.

[0091] Through the above steps, the power consuming equipment morpheme library is completed. After that, when the power consuming equipment shows the new power consuming behavior mode again, the edge computing unit 10 can accurately identify it as a new power consuming equipment morpheme by using the updated model and report it to the cloud server 20, thereby ensuring the self-adaptability of the entire big data processing system to new equipment or new working conditions of the equipment and the accuracy of long-term operation.

[0092] Referring to the accompanying Figure 7 , Figure 7 The flow chart of the big data processing method for socket energy efficiency optimization and energy consumption prediction according to one embodiment of the application. The method provided by the application applied to the system of the above embodiment one can include the following steps:

[0093] S10, real-time collection of original electric signals of power consuming equipment connected with the socket, and processing the original electric signals into power consuming equipment morpheme sequences arranged in time sequence and representing the basic power consuming behavior of the power consuming equipment. This step is executed by the electrical appliance morpheme extraction module 110 in the edge computing unit 10. Specifically, this step includes: collecting original electric signals at a preset frequency to form a signal window; performing time-frequency domain transformation on the signal window to generate a time-frequency spectrum; using a pre-installed convolutional neural network to classify the time-frequency spectrum to map each signal window to a power consuming equipment morpheme, thereby generating a power consuming equipment morpheme sequence.

[0094] S20, receiving the power consuming equipment morpheme sequence, and predicting the power consuming behavior in a future period of time based on the historical context of the power consuming equipment morpheme sequence, and generating the corresponding future power consuming equipment morpheme sequence, the energy consumption trend prediction result, and the attention weight representing the dependency relationship between the historical power consuming equipment morphemes. This step is executed by the sequence prediction module 210 in the cloud server 20. Specifically, this step uses an encoder-decoder model based on the Transformer architecture, the encoder part of which processes the historical context, and the decoder part of which generates the future power consuming equipment morpheme sequence. At the same time, the attention weight is calculated through the multi-head self-attention mechanism built in the model, and the energy consumption trend prediction result is generated by combining the future sequence with the pre-installed power consumption mapping relationship.

[0095] S30, combining the energy consumption trend prediction result and the attention weight to generate a comprehensive state vector including the power consuming equipment behavior context information. This step is executed by the decision state generation module 220 in the cloud server 20. Specifically, this step includes: performing attribution analysis on the historical power consuming equipment morphemes based on the attention weight to identify key historical power consuming equipment morphemes and generate a context criticality vector; splicing the context criticality vector with the energy consumption trend prediction result, and optionally splicing with the external environment information vector and the user preference vector to generate a comprehensive state vector.

[0096] S40, receiving the comprehensive state vector and outputting the energy efficiency optimization intervention instruction for the socket based on the comprehensive state vector. This step is performed by the active intervention decision module 230 in the cloud server 20. Specifically, this step uses a reinforcement learning agent as a pre-trained deep Q network model, inputs the comprehensive state vector as the current state, and selects an action that can maximize the long-term cumulative reward based on a multi-objective reward function (which at least evaluates the expected energy saving amount, user satisfaction and device health) by the reinforcement learning agent to generate the energy efficiency optimization intervention instruction.

[0097] In a further embodiment, the method can further comprise the following steps:

[0098] S50, when a new power consumption behavior pattern that cannot be accurately described by the existing power consumption device morpheme is identified, generating a new power consumption device morpheme definition and issuing the new power consumption device morpheme definition to the appliance morpheme extraction module to dynamically expand the power consumption device morpheme library. This step is performed by the dynamic morpheme library updating module 240 in the cloud server 20 to ensure the adaptive ability of the system to new devices and new working conditions.

[0099] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A big data processing system for socket energy efficiency optimization and energy consumption prediction, characterized in that, The utility model relates to an energy consumption prediction method and device based on edge computing and cloud computing, and a power socket. The power socket is provided with an edge computing unit, which includes an appliance morpheme extraction module. The appliance morpheme extraction module is used to collect raw power signals of an electrical equipment connected to the power socket in real time, and process the raw power signals into a sequence of appliance morphemes arranged in time sequence, which represent the basic power consumption behavior of the electrical equipment. A cloud server is communicatively connected to the edge computing unit, and includes a sequence prediction module, a decision state generation module, and an active intervention decision module. The sequence prediction module is used to receive the sequence of appliance morphemes, and based on the historical context of the sequence of appliance morphemes, predict the power consumption behavior in a future period of time, and generate a corresponding future sequence of appliance morphemes, an energy consumption trend prediction result, and attention weights representing the dependency relationship between historical appliance morphemes. The decision state generation module is used to generate a comprehensive state vector including the context information of the electrical equipment behavior by combining the energy consumption trend prediction result and the attention weights. The active intervention decision module is used to receive the comprehensive state vector, and output an energy efficiency optimization intervention instruction for the power socket based on the comprehensive state vector. The appliance morpheme extraction module is specifically used to: Collect the raw power signals at a preset frequency to form a signal window. Perform time-frequency domain transformation on the signal window to generate a time-frequency spectrum. Analyze and classify the time-frequency spectrum using a pre-installed neural network to map each signal window to an appliance morpheme, thereby generating the sequence of appliance morphemes arranged in time sequence, which represent the basic power consumption behavior of the electrical equipment.

2. The big data processing system for socket energy efficiency optimization and energy consumption prediction of claim 1, wherein, The sequence prediction module is specifically used to: Process the historical context of the sequence of appliance morphemes using an encoder-decoder model based on the Transformer architecture for use by the encoder part, and generate the future sequence of appliance morphemes by the decoder part. The model further combines the future sequence of appliance morphemes with a pre-installed mapping relationship between appliance morphemes and power consumption to generate the energy consumption trend prediction result. The attention weights are calculated through the multi-head self-attention mechanism built in the encoder-decoder model. The historical context of the sequence of appliance morphemes is a sub-sequence of appliance morphemes of a preset length before the time point of prediction.

3. The big data processing system for socket energy efficiency optimization and energy consumption prediction of claim 1, wherein, The decision state generation module is specifically used to: Based on the attention weights, perform attribution analysis on the historical appliance morphemes in the sequence of appliance morphemes to identify key historical appliance morphemes that have a decisive effect on the current state. According to the result of the attribution analysis, generate a context criticality vector that quantifies the influence degree of the key historical appliance morphemes. Concatenate the context criticality vector with the prediction result to generate the comprehensive state vector including the context information of the electrical equipment behavior.

4. The big data processing system for socket energy efficiency optimization and energy consumption prediction of claim 3, wherein, The decision state generation module is further configured to splice an external environment information vector and a user preference vector when generating the comprehensive state vector, the external environment information vector including at least one of real-time electricity price information and power grid load state, and the user preference vector including at least one of historical intervention instruction acceptance rate and user preset mode preference.

5. The big data processing system for socket energy efficiency optimization and energy consumption prediction of claim 1, wherein, The active intervention decision module is specifically configured to: The active intervention decision module is specifically configured to:

6. The big data processing system for socket energy efficiency optimization and energy consumption prediction of claim 5, wherein, The active intervention decision module is specifically configured to:

7. The big data processing system for socket energy efficiency optimization and energy consumption prediction of claim 1, wherein, The active intervention decision module is specifically configured to: The cloud server further includes: The dynamic morpheme library updating module is configured to generate a new definition of a power consumption equipment morpheme when the sequence prediction module identifies a new power consumption behavior mode that cannot be accurately described by an existing power consumption equipment morpheme, and to distribute the new definition of the power consumption equipment morpheme to the appliance morpheme extraction module to dynamically expand the power consumption equipment morpheme library.

Images