A management method, system, device and medium for energy conservation prediction

Abstract

The present invention discloses a management method, system, device and medium for energy conservation prediction, relating to the technical field of energy conservation optimization, including collecting multi-source data of energy usage terminals and performing preprocessing. Feature extraction and data analysis are carried out on the preprocessed data, and energy consumption prediction is carried out based on a deep neural network model. An energy conservation strategy is generated according to the prediction result, and dynamic optimization adjustment is carried out through a feedback mechanism. The method of the present invention adaptively adjusts the feature weights through L-VAE, reduces the model calculation complexity, improves the feature representation ability, enables the energy consumption prediction to still operate efficiently in the case of large data volume and complex features, and avoids information loss. The adaptability of the energy consumption prediction is improved, so that in the face of sudden load changes, the prediction result can be quickly adjusted to ensure the prediction accuracy, thereby supporting the formulation of more refined energy conservation strategies.

Description

Technical Field

[0001] The present invention relates to the technical field of energy saving optimization, and in particular to a management method, system, equipment and medium for energy saving prediction. Background Art

[0002] Energy conservation management has become a key research area in industry, commerce, and public infrastructure. Traditional energy management methods primarily rely on rule-based policy formulation, such as fixed-time scheduling and device energy consumption threshold control. However, these approaches struggle to adapt to complex energy demand fluctuations. In recent years, data-driven energy forecasting technologies have rapidly developed, particularly energy consumption forecasting models based on machine learning and deep learning. These models analyze future energy consumption trends using historical data, improving the intelligence of energy conservation management. Furthermore, the application of reinforcement learning in energy optimization decision-making has also gained increasing attention. By constructing a state-action-reward mechanism, it can dynamically optimize energy allocation strategies and improve energy efficiency. However, existing technologies still suffer from significant data noise, inadequate feature extraction, and difficulty predicting short-term load fluctuations, limiting the real-time and accuracy of energy conservation strategies. Therefore, efficiently and accurately predicting energy consumption and optimizing energy conservation strategies based on the predicted results have become major technical challenges in the field of energy management.

[0003] Currently, energy consumption forecasting and energy-saving optimization mainly rely on traditional time series analysis methods (such as ARIMA) and neural network-based forecasting models (such as LSTM and GRU). However, these methods have the following limitations when faced with the dynamic characteristics of complex energy systems:

[0004] Insufficient feature selection and dimensionality reduction: Existing methods usually use principal component analysis (PCA) or simple statistical screening methods to extract features from energy consumption data, but it is difficult to capture the correlation between nonlinear features, resulting in the model input data being too high in dimension, affecting prediction accuracy.

[0005] Limited ability to predict short-term energy consumption fluctuations: Traditional deep learning methods, such as LSTM or GRU, although they have certain advantages in time series modeling, have a lag effect when dealing with sharp fluctuations in short-term energy consumption, resulting in large prediction errors.

[0006] Energy consumption optimization strategies lack adaptability: Existing energy-saving optimization methods mostly rely on fixed rules or simple optimization methods, making it difficult to adjust strategies in real time and unable to adapt to environmental changes (such as load changes, weather impacts, etc.), resulting in poor execution of energy-saving strategies.

[0007] Imperfect feedback mechanism: Many existing systems lack closed-loop feedback mechanisms. After energy-saving strategies are implemented, they cannot be dynamically optimized based on actual energy consumption. This makes it difficult to continuously improve the strategy optimization process and reduces long-term energy-saving effects. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed.

[0009] Therefore, the technical problems solved by the present invention are as follows: existing energy consumption prediction and management methods have problems such as insufficient feature selection, limited short-term energy consumption fluctuation prediction ability, lack of self-adaptability of energy-saving strategies, imperfect feedback mechanisms, etc., and the problem of how to construct a closed-loop energy-saving management system to achieve accurate energy consumption prediction and dynamic optimization of energy-saving control.

[0010] To solve the above technical problems, the present invention provides the following technical solution: a management method for energy-saving prediction, including: collecting multi-source data of energy-using terminals and performing preprocessing. Performing feature extraction and data analysis on the preprocessed data, and performing energy consumption prediction based on a deep neural network model. Generating an energy-saving strategy according to the prediction result and performing dynamic optimization and adjustment through a feedback mechanism. The deep neural network model includes an input layer, a variational autoencoder layer, a time series modeling layer, a Transformer prediction layer, a fully connected layer, and an output layer. The input layer receives a feature set that affects the energy consumption prediction of the park, and embeds it into a high-dimensional feature space through a multi-head self-attention mechanism. The variational autoencoder layer introduces a Laplace variational autoencoder to perform feature dimensionality reduction on the input park energy consumption data, removes redundant information, optimizes data representation, regularizes the latent variables output by the encoder using a Laplace prior distribution, optimizes the variational lower bound through Laplace variational inference, and generates a low-dimensional interpretable feature representation by balancing the reconstruction error of the decoder and the KL divergence. The time series modeling layer models the short-term changes of energy consumption data through a gated recurrent unit, and dynamically adjusts the hidden state in combination with an attention mechanism. The Transformer prediction layer adopts a local spatio-temporal attention mechanism, adds local time window regulation to the global Transformer architecture, dynamically divides the input sequence based on a sliding time window, applies local attention weights to the time steps within the window, adjusts the contribution ratio of global and local attention through a time weighting mechanism, and optimizes the joint modeling ability of short-term energy consumption fluctuations and long-term trends. The fully connected layer processes the temporal features output by the Transformer layer and uses the ReLU activation function to enhance the non-linear mapping ability. The output layer converts the output of the fully connected layer into the final energy consumption prediction result.

[0011] As a preferred solution of the management method for energy-saving prediction according to the present invention, wherein: the preprocessing includes: collecting park energy consumption data through energy-using terminals and performing time series alignment on the data based on a time synchronization mechanism. Cleaning the park energy consumption data after time series alignment, including removing duplicate data, abnormal data, and filling in missing data.

[0012] As a preferred solution of the management method for energy conservation prediction described in the present invention, wherein: the feature extraction and data analysis include: segmenting the preprocessed campus energy consumption data by time series, adopting a feature selection method based on multi-layer nested clustering, classifying and summarizing the global pattern and local pattern of the campus energy consumption data respectively, and calculating the mutual influence degree between feature variables by combining mutual information analysis, so as to screen out the feature set that has the most significant influence on the campus energy consumption prediction.

[0013] As a preferred solution of the management method for energy conservation prediction described in the present invention, wherein: the deep neural network model includes: the input layer receives the feature set that has the most significant influence on the campus energy consumption prediction, and adopts a multi-head self-attention embedding mechanism to map multi-dimensional input data to a high-dimensional feature space. The variational autoencoder layer introduces a Laplace variational autoencoder to perform feature dimensionality reduction on the input campus energy consumption data, removes redundant information, optimizes data representation, regularizes the latent variables output by the encoder by using a Laplace prior distribution, optimizes the variational lower bound through Laplace variational inference, optimizes the modeling ability for the sparsity and mutation characteristics of the energy consumption data, and combines the reconstruction error of the decoder and the KL divergence balance to generate a low-dimensional interpretable feature representation. The time series modeling layer uses a gated recurrent unit for time series modeling to learn the short-term and long-term energy consumption change trends from the input time series energy consumption data. The hidden state weights of the gated recurrent unit are dynamically adjusted through an attention mechanism. The Transformer prediction layer adopts a local spatio-temporal attention mechanism, adds local time window regulation on the global Transformer architecture, dynamically divides the input sequence based on a sliding time window, applies local attention weights to the time steps within the window, and adjusts the contribution ratio of global and local attention through a time weighting mechanism, so as to optimize the joint modeling ability for short-term energy consumption fluctuations and long-term trends. The fully connected layer receives the time series features output by the Transformer prediction layer and performs a non-linear transformation, and uses a ReLU activation function to enhance the non-linear fitting ability. The output layer converts the output of the fully connected layer into the final energy consumption prediction result.

[0014] As a preferred solution of the energy-saving prediction management method described in the present invention, wherein: the energy consumption prediction includes: standardizing the feature set that has the most significant impact on the park energy consumption prediction, and converting it to the same scale using the Z-score normalization method. Using an embedding layer based on multi-head self-attention to input the feature set data of the same scale into a deep neural network to capture the long-term and short-term dependencies between different features. Based on the variational autoencoder for feature dimensionality reduction, converting the features into low-dimensional latent vector representations, and introducing a gated recurrent unit (GRU) for time series modeling to extract the temporal features of the energy consumption changes. Inputting the output result of the GRU into a fully connected layer, and using the Softmax activation function to calculate the probability distribution of the park energy consumption at different time steps, and combining with the Bayesian regression model calculation to obtain the final park energy consumption prediction result.

[0015] As a preferred solution of the energy-saving prediction management method described in the present invention, wherein: generating an energy-saving strategy based on the prediction result includes: calculating and generating an energy-saving strategy for the park energy consumption according to the park energy consumption prediction result, using a reinforcement learning algorithm to evaluate the adaptability of different strategies, and screening the optimal energy-saving strategy.

[0016] As a preferred solution of the energy-saving prediction management method described in the present invention, wherein: and performing dynamic optimization and adjustment through a feedback mechanism includes: implementing the optimal energy-saving strategy, continuously monitoring the actual energy consumption data, obtaining the real-time device operation status, environmental parameters, and load levels, and forming a feedback data set of the execution effect of the energy-saving strategy. Based on the feedback data set, using an adaptive error analysis model to calculate the error between the actual energy consumption value and the park energy consumption prediction result, and analyzing the error source. According to the error analysis result, dynamically optimizing the energy-saving strategy.

[0017] Another object of the present invention is to provide an energy-saving prediction management system, which can perform feature extraction and data analysis on the preprocessed data, and perform energy consumption prediction based on a deep neural network model, solving the problem that traditional deep learning methods (LSTM or GRU) have a lag effect when dealing with drastic short-term energy consumption fluctuations, resulting in large prediction errors.

[0018] As a preferred solution of the energy-saving prediction management system described in the present invention, wherein: it includes a preprocessing module, a prediction module, and an optimization module.

[0019] The preprocessing module is used to collect multi-source data of energy usage terminals and perform preprocessing.

[0020] The prediction module is used to perform feature extraction and data analysis on the preprocessed data, and perform energy consumption prediction based on a deep neural network model.

[0021] The optimization module is used to generate an energy-saving strategy based on the prediction result and perform dynamic optimization and adjustment through a feedback mechanism.

[0022] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the management method for energy-saving prediction are implemented.

[0023] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the management method for energy-saving prediction are implemented.

[0024] The beneficial effects of the present invention: The management method for energy-saving prediction provided by the present invention uses energy consumption terminals in the park to collect energy consumption data, performs time series alignment on the data through a time synchronization mechanism, and simultaneously performs data cleaning, including removing duplicate data and abnormal data, and filling missing data using the interpolation method. This ensures the quality of the subsequent model training data, enables the energy consumption prediction model to learn on a stable and clean data set, improves the prediction accuracy, and reduces errors.

[0025] Segment the preprocessed energy consumption data in the park by time series, adopt a feature selection method based on multi-layer nested clustering, classify and summarize the global pattern and the local pattern respectively, and calculate the mutual influence degree between feature variables by combining mutual information analysis, and screen out the feature set that has the most significant impact on energy consumption prediction. This step makes the extraction of energy consumption characteristics more targeted by classifying and summarizing the global and local patterns, and at the same time accurately screens key variables with the help of mutual information analysis, avoiding the influence of redundant features on the prediction accuracy. It improves the interpretability of energy consumption data, reduces the noise brought by irrelevant features, enables the prediction model to learn the energy consumption law more efficiently, improves the generalization ability, and can still maintain a high prediction accuracy in different park environments.

[0026] The variational autoencoder layer introduces the Laplace variational autoencoder (L-VAE) to automatically capture the structural information of high-dimensional data and uses Laplace prior regularization encoding to make the low-dimensional representation more interpretable. This step adaptively adjusts the feature weights through L-VAE, making the dimensionality reduction process of energy consumption data more accurate, which can not only reduce data redundancy but also retain the most critical information for prediction. It reduces the computational complexity of the model, improves the feature representation ability, enables the energy consumption prediction to still operate efficiently in the case of large data volume and complex features, and avoids information loss.

[0027] The time series modeling layer uses a gated recurrent unit (GRU) for time series modeling, learns the short-term and long-term energy consumption change trends from the input time series energy consumption data, and dynamically adjusts the hidden state weights of the GRU through an attention mechanism. The Transformer prediction layer introduces a local spatio-temporal attention mechanism (LSTA). Based on the global self-attention calculation, it combines the adjustment of local time window weights to enhance the modeling ability for short-term fluctuations. This step combines the long-term dependence ability of the GRU and the global modeling ability of the Transformer, and enhances the ability to capture short-term energy consumption changes through the LSTA, enabling energy consumption prediction to consider both long-term trends and make a rapid response to short-term load changes. It improves the adaptability of energy consumption prediction, enabling the system to quickly adjust the prediction results in the face of sudden load changes, ensuring the prediction accuracy, and thus supporting the formulation of more refined energy-saving strategies.

[0028] The present invention constructs a complete closed-loop energy-saving prediction management method. Compared with traditional methods, the invention of our side has significant advantages in data dimensionality reduction, short-term energy consumption prediction, energy-saving strategy optimization and dynamic adjustment, making energy consumption prediction more accurate and energy-saving strategies more efficient. Brief Description of the Drawings

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0030] Figure 1 It is the overall flowchart of a management method for energy-saving prediction provided by the first embodiment of the present invention. Detailed Embodiments

[0031] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will make a detailed description of the specific embodiments of the present invention in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0032] Example 1, referring to Figure 1 , which is an embodiment of the present invention, provides a management method for energy-saving prediction, including:

[0033] S1: Collect multi-source data of energy usage terminals and perform preprocessing.

[0034] Collect the energy consumption data of the park through the energy - using terminals, and align the data in time series based on the time - synchronization mechanism. Clean the energy consumption data of the park after time - series alignment, including removing duplicate data, abnormal data, and filling in missing data.

[0035] It should be noted that a preferred solution for collecting the energy consumption data of the park through the energy - using terminals specifically includes that the system collects power consumption data, environmental parameter data, load and equipment status data, and user behavior data through the energy - using terminals (including smart meters, smart air - conditioner controllers, lighting management systems, temperature and humidity sensors, etc.) installed in each energy - using unit of the park. The power consumption data includes total active power, voltage, current, power factor, and sub - circuit energy consumption data. The environmental parameter data includes outdoor environmental temperature, humidity, solar radiation intensity, indoor temperature and humidity. The load and equipment status data includes equipment start - stop status, running time, and load change rate. The user behavior data includes the flow data of office staff and the working day / holiday mark.

[0036] Furthermore, a preferred solution for aligning the data in time series based on the time - synchronization mechanism specifically includes that since the data comes from multiple terminal devices, there may be clock deviations in their collection times. Therefore, the Network Time Protocol (NTP) is used for time synchronization to ensure that all data is stored and calculated based on the same time reference. Linear interpolation is used to align the data collected at different times to a fixed time interval (such as one data point every 5 minutes). For some data points with discontinuous time series, local weighted regression (LOWESS) is used for smoothing to reduce the impact of mutation points.

[0037] Even further, a preferred solution for removing duplicate data, abnormal data, and filling in missing data specifically includes removing duplicate data by de - duplicating based on the device ID and timestamp to ensure that there is only one piece of data at each time point. The Z - score method (with a threshold set to 3σ) is used for outlier detection and elimination to detect abnormal - fluctuation power data, such as suddenly emerging extreme high - load situations, and eliminate the outlier points exceeding 3 times the standard deviation. Numerically impossible values are directly eliminated. For missing values in data collection, for short - term missing values (<1 hour), linear interpolation is used for filling to ensure the smoothness of the data in a short period. For long - term missing values (>1 hour), LSTM time - series prediction is used to fill in the long - term data missing to avoid the impact of data fluctuations on model training.

[0038] It should be noted that the core goal of S1 is to obtain high-quality park energy consumption data to ensure the accuracy of subsequent energy consumption prediction and the rationality of energy-saving strategies. The design idea is to improve the integrity, consistency, and timeliness of data through multi-source data collection and preprocessing methods such as time synchronization, data cleaning, anomaly detection, and missing value filling. The effect of such design is to eliminate data errors, improve data availability, and make the data input into the energy consumption prediction model more representative. For example, time synchronization can ensure that data from different devices are aligned under the same time reference, data cleaning can remove invalid or abnormal data, and missing value filling can reduce the bias in model training. And the reason for adopting this design is that the energy usage in the park is time-varying, complex, and diverse, the energy consumption characteristics of different types of devices are different, and the data may be affected by factors such as sensor failures, network fluctuations, and equipment aging. Without effective preprocessing, it will lead to too much noise in the data input into the prediction model, affecting the accuracy of energy consumption prediction. Therefore, reasonable data collection and preprocessing are the key links to ensure the effectiveness of subsequent prediction and optimization decisions.

[0039] S2: Extract features and perform data analysis on the preprocessed data, and conduct energy consumption prediction based on a deep neural network model.

[0040] Segment the time series of the preprocessed park energy consumption data, adopt a feature selection method based on multi-layer nested clustering, classify and summarize the global pattern and local pattern of the park energy consumption data respectively, and calculate the mutual influence degree between feature variables in combination with mutual information analysis to screen the feature set that has the most significant impact on park energy consumption prediction.

[0041] A preferred solution for classifying and summarizing the global pattern and local pattern of park energy consumption data specifically includes, in the datasets of each time window, using hierarchical clustering to divide the global pattern respectively, and calculating data similarity using Euclidean distance. On the basis of hierarchical clustering, use nested density clustering to further refine the classification of local time windows to ensure that sudden events such as temporary start and stop of equipment and sudden increase in load can be accurately captured in the short-term energy consumption characteristics.

[0042] A preferred solution for screening the feature set that has the most significant impact on park energy consumption prediction specifically includes calculating the dependence relationship between different energy consumption variables using mutual information entropy. Set the information threshold to 0.05. If the mutual information between a certain feature and the target variable (future energy consumption) is lower than this threshold, then eliminate this feature. According to the calculation results of mutual information, select the top 10 variables with the highest mutual information values as core features, and use principal component analysis for further dimensionality reduction to screen the principal components that can explain more than 90% of the variance as the final feature input into the deep neural network model.

[0043] The deep neural network model includes an input layer that receives the feature set with the most significant impact on the prediction of campus energy consumption. The multi-head self-attention embedding mechanism is used to map multi-dimensional input data to a high-dimensional feature space. The variational autoencoder layer introduces a Laplace variational autoencoder to perform feature dimensionality reduction on the input campus energy consumption data, remove redundant information, optimize data representation, regularize the latent variables output by the encoder using a Laplace prior distribution, optimize the variational lower bound through Laplace variational inference, enhance the modeling ability for the sparsity and mutation characteristics of energy consumption data, and balance the reconstruction error of the decoder and the KL divergence to generate low-dimensional interpretable feature representations. The time series modeling layer uses a gated recurrent unit for time series modeling to learn the short-term and long-term energy consumption change trends from the input time series energy consumption data. The hidden state weights of the gated recurrent unit are dynamically adjusted through an attention mechanism. The Transformer prediction layer uses a local spatio-temporal attention mechanism, adds local time window regulation to the global Transformer architecture, dynamically divides the input sequence based on a sliding time window, applies local attention weights to the time steps within the window, and adjusts the contribution ratio of global and local attention through a time weighting mechanism to enhance the joint modeling ability for short-term energy consumption fluctuations and long-term trends. The fully connected layer receives the time series features output by the Transformer prediction layer and performs a non-linear transformation, using the ReLU activation function to enhance the non-linear fitting ability. The output layer converts the output of the fully connected layer into the final energy consumption prediction result.

[0044] Furthermore, a preferred solution of the variational autoencoder layer specifically includes that in the variational autoencoder (VAE) layer, a Laplace variational autoencoder (L-VAE) is introduced to enhance the robustness in the data dimensionality reduction process and reduce non-linear noise interference. The standard VAE uses a Gaussian distribution as the prior distribution, which is difficult to effectively model sparse data and energy consumption mutation data. However, the invention of our side uses a Laplace prior distribution (LaplacePrior), which is more suitable for the sparse characteristics of energy consumption data and can more effectively remove irrelevant information. It is expressed as:

[0045] ;

[0046] Wherein, represents the loss function of the Laplace variational autoencoder. represents the mathematical expectation under the variational distribution and is used to calculate the average log-likelihood value of the reconstructed data. represents the input high-dimensional energy consumption feature matrix. represents the low-dimensional latent variable. represents the variational distribution. represents the decoder generation distribution. represents the Kullback-Leibler divergence, which is used to measure the similarity between the variational distribution and the prior distribution. represents the reconstruction loss term, which is used to measure the input data and the reconstructed by the latent variable represents measuring the distribution of the latent variable and the prior distribution to make the features after dimensionality reduction more sparse and stable. represents the weight of the KL divergence, with a value , which is used to balance the effects of reconstruction error and prior matching. Adopt the Laplace distribution:

[0047] ;

[0048] where, represents the mean, represents the scale parameter, taking to maintain data sparsity.

[0049] Furthermore, a preferred solution of the Transformer prediction layer specifically includes introducing a local spatio-temporal attention mechanism (Local Spatio-Temporal Attention, LSTA) in the Transformer prediction layer to enhance the model's sensitivity to short-term fluctuations while maintaining the long-term trend modeling ability. The standard Transformer uses global self-attention calculation, ignoring the contribution of local time windows to the prediction results, resulting in difficulty in modeling short-term sudden loads. And the computational complexity is relatively high, consuming too much computing resources in the energy consumption prediction scenario. Adopt local window selection (Sliding Time Window Selection) to dynamically adjust the window size and improve the modeling ability for short-term load changes. Combine temporal weighted attention (Temporal Weighted Attention) to enhance the adaptability of the Transformer to different time steps. It is expressed as:

[0050] ;

[0051] where, represents the attention weight within the local time window. represents calculating the correlation score of the local time window. represents calculating the correlation score of the local time window at time step ​ represents the time window size and is set to ensure that the model can adapt to the requirements of different time scales. represents the mapping matrix of the query, and represents the mapping matrix of the key. represents the time step of the query vector, represents the time step of the key vector to ensure the correlation calculation of time series data. represents the transpose of the weight matrix. .8, then the short-term burst load has a greater impact on the prediction. If .2, then it mainly relies on long-term trend modeling.

[0052] The energy consumption prediction steps are as follows:

[0053] Standardize the feature set that has the most significant impact on the campus energy consumption prediction, and convert it to the same scale using the Z-score normalization method. Input the data of the feature set of the same scale into the deep neural network using the embedding layer based on multi-head self-attention to capture the long-term and short-term dependencies between different features. Perform feature dimensionality reduction based on the variational autoencoder, convert the features into low-dimensional latent vector representations, and introduce the gated recurrent unit GRU for time series modeling to extract the temporal features of the energy consumption changes. Input the output result of the GRU into the fully connected layer, and use the Softmax activation function to calculate the probability distribution of the campus energy consumption at different time steps, and combine the Bayesian regression model calculation to obtain the final campus energy consumption prediction result.

[0054] It should be noted that the core goal of S2 is to extract key features from the preprocessed energy consumption data and perform energy consumption prediction through a deep neural network model to ensure the accuracy and generalization ability of the prediction. The design idea is to use multi-layer nested clustering, mutual information analysis, and principal component analysis to screen the most important features, and combine with Laplace variational autoencoder (L-VAE) for dimensionality reduction to remove redundant information. Then, a local spatio-temporal attention mechanism (LSTA) is used to optimize the Transformer to improve the modeling ability of short-term load fluctuations, so as to enhance the accuracy and adaptability of time series prediction. This step ensures that the feature data input into the deep learning model is the most representative through multi-layer clustering and mutual information analysis, and avoids the influence of redundant data on the prediction results. L-VAE is used to reduce the dimensionality of energy consumption features, making the data representation more sparse and improving the stability of model training. By optimizing the Transformer with LSTA, the model can take into account both short-term fluctuations and long-term trends, improving the accuracy of energy consumption prediction. Because traditional feature selection methods (such as PCA) cannot effectively distinguish short-term abnormal energy consumption fluctuations, while multi-layer nested clustering can refine short-term and long-term patterns and improve the adaptability of feature extraction. The Gaussian prior distribution of traditional VAE is difficult to effectively model the sparse characteristics of energy consumption data, while L-VAE can enhance the feature sparsification ability and improve the dimensionality reduction effect. Traditional Transformer has a high computational complexity and is difficult to handle short-term energy consumption fluctuations. LSTA adds a local time window on the basis of global attention, enabling the model to more effectively learn the impact of sudden loads and improving the robustness of prediction.

[0055] S3: Generate an energy-saving strategy based on the prediction result and perform dynamic optimization and adjustment through a feedback mechanism.

[0056] Calculate and generate the energy-saving strategy for the park's energy consumption according to the park's energy consumption prediction result, use the reinforcement learning algorithm to evaluate the adaptability of different strategies, and screen the optimal energy-saving strategy. Perform dynamic optimization and adjustment through a feedback mechanism. First, execute the optimal energy-saving strategy, continuously monitor the actual energy consumption data, obtain the real-time device operation status, environmental parameters, and load level, and form a feedback data set of the execution effect of the energy-saving strategy. Based on the feedback data set, use an adaptive error analysis model to calculate the error between the actual energy consumption value and the park's energy consumption prediction result, and analyze the error source. According to the error analysis result, dynamically optimize the energy-saving strategy.

[0057] A preferred solution for evaluating the adaptability of different strategies using a reinforcement learning algorithm and screening the optimal energy-saving strategy specifically includes the following steps. In this step, an energy-saving strategy for the campus energy consumption is calculated and generated based on the campus energy consumption prediction results, and a reinforcement learning algorithm is used to evaluate the adaptability of different strategies to screen the optimal energy-saving strategy. First, based on the energy consumption prediction results, a multi-objective optimization model is established. The optimization objectives include minimizing the energy consumption cost, reducing the peak load, and improving the equipment utilization rate. The constraint conditions cover the equipment operation time, load balance, and user comfort parameters. For this optimization problem, an intelligent decision-making framework based on reinforcement learning is adopted, including a state space, an action space, a reward function, policy evaluation, and policy selection. The state space contains information such as the current operation state of campus equipment, load level, and environmental factors (temperature, humidity). The action space defines possible energy-saving operations, such as adjusting the equipment operation time, optimizing the load distribution, and reducing the power consumption of unnecessary equipment. The reward function combines three factors: energy-saving effect, user comfort, and equipment life, calculates the benefits of different strategies, and uses a weighted objective function to balance multiple optimization objectives. Policy evaluation uses a deep Q-network (DQN) for reinforcement learning training and combines an experience replay mechanism to enable the policy to adapt to the changes in campus energy consumption. Policy selection is based on the results of multiple iterative trainings, and a Monte Carlo tree search (MCTS) is used to screen the optimal energy-saving strategy to ensure that the policy has generalization ability in different environments. After implementing the optimal energy-saving strategy, dynamic optimization and adjustment are carried out through a feedback mechanism, that is, the actual energy consumption data is continuously monitored, and the equipment operation state, environmental parameters, and load level are collected to form a feedback data set of the implementation effect of the energy-saving strategy. Subsequently, an adaptive error analysis model is used to calculate the error between the actual energy consumption value and the predicted energy consumption value. The error sources include:

[0058] Load fluctuation error: The energy consumption deviation caused by the temporary start and stop of equipment, production plan adjustment, etc.

[0059] Environmental impact error: The error in energy consumption demand affected by factors such as weather changes and temperature and humidity changes.

[0060] Equipment operation error: The power loss deviation caused by factors such as equipment aging and untimely maintenance.

[0061] A preferred solution for dynamically optimizing the energy-saving strategy based on the error analysis results specifically includes using Bayesian optimization to adjust the reward function in the reinforcement learning model, enabling the strategy to adapt to the energy consumption patterns of different scenarios. Combining an adaptive sliding window to optimize the input data of the energy consumption prediction model and enhancing the model's response ability to short-term changes. Using multi-layer clustering analysis to reclassify different device loads and adjusting the energy-saving strategy for different categories of devices. For example, a more conservative energy-saving strategy is adopted for high-fluctuation load devices, and a more aggressive optimization strategy is adopted for stable load devices. Finally, the optimized energy-saving strategy will be re-executed in the next energy-saving cycle, and the feedback optimization process will be repeated to form a closed-loop energy-saving management system, enabling the energy-saving strategy to continuously adapt to the dynamically changing energy consumption demand and improving the overall energy utilization efficiency of the park.

[0062] It should be noted that the core objective of S3 is to formulate an energy-saving strategy based on the energy consumption prediction results and dynamically optimize it through reinforcement learning and feedback mechanism, enabling the energy-saving strategy to continuously adapt to the changing energy consumption demand. The design idea is based on a multi-objective optimization model, comprehensively considering energy consumption cost, peak load, and equipment utilization rate, using a reinforcement learning algorithm (DQN+MCTS) to screen the optimal energy-saving strategy, and realizing the adaptive optimization of the strategy through the feedback mechanism and error analysis. This step automatically finds the optimal strategy through reinforcement learning to ensure that the energy-saving solution can operate efficiently under different load conditions. Using the feedback mechanism to continuously adjust the strategy enables the energy-saving solution to adapt to sudden load fluctuations or environmental changes and improve energy utilization efficiency. Error analysis can accurately identify the deviation in strategy execution, forming a closed-loop optimization for energy-saving management and improving the long-term energy-saving effect. Since the traditional rule optimization method has poor adaptability in scenarios with large load fluctuations, reinforcement learning can dynamically adjust the strategy to improve the flexibility of the energy-saving solution. Factors such as equipment aging and environmental changes may affect the strategy execution effect, so a feedback mechanism is needed for adaptive optimization. By adjusting the reward function through Bayesian optimization, the strategy can be prevented from converging to a local optimum, improving the generalization ability of energy-saving management and making the energy consumption management of the entire park more intelligent and efficient.

[0063] Example 2, an embodiment of the present invention, provides a management method for energy-saving prediction. To verify the beneficial effects of the present invention, scientific demonstrations are carried out through economic benefit calculations and simulation experiments.

[0064] The experiment was carried out in a large industrial park, which has multiple high-energy-consuming devices, such as central air-conditioning systems, intelligent lighting systems, electric mechanical equipment, etc. The experiment used real device energy consumption data for testing and combined the multi-layer nested clustering feature selection method, Laplace variational autoencoder (L-VAE) dimensionality reduction, local spatio-temporal attention (LSTA) optimized Transformer prediction model, and energy-saving strategy optimization method based on reinforcement learning (DQN+MCTS) proposed by the invention to carry out energy consumption prediction and energy-saving control strategy formulation.

[0065] In the experiment, energy consumption data of each area was first obtained through smart meters, temperature and humidity sensors, load monitoring devices, etc., and the core parameters were recorded. All data was aligned using a time synchronization mechanism and underwent data cleaning to remove outliers to ensure data integrity.

[0066] A multi-layer nested clustering method was used to analyze the global and local energy consumption patterns, and the influence degree of key features on energy consumption was calculated through mutual information analysis, so as to screen out the most representative feature set. Then, the L-VAE model was used to reduce the dimension of the features to remove redundant information and improve the data processing efficiency. Subsequently, a Transformer prediction model combining GRU and LSTA was used to predict the energy consumption trend, where GRU was responsible for learning the long-term energy consumption change trend and LSTA improved the sensitivity to short-term fluctuations. The predicted energy consumption data by the model will be directly used for the formulation of energy-saving strategies.

[0067] Based on the energy consumption prediction results, reinforcement learning optimization (DQN + MCTS) was used to formulate energy-saving strategies, and the optimization objectives included reducing peak load, improving equipment utilization rate, and reducing the power consumption of unnecessary equipment. The system adaptively adjusted the strategy parameters, enabling the energy-saving strategy to adapt to the dynamically changing load environment in real time.

[0068] After the implementation of the energy-saving strategy, the system continuously monitored the energy consumption data of the equipment, compared the deviation between the actual energy consumption and the predicted energy consumption, and calculated the implementation effect of the strategy. An adaptive error analysis model was used to calculate the error source, and the strategy parameters were adjusted based on Bayesian optimization to ensure the continuous optimization of energy-saving management. The experiment was conducted for comparison in two time periods of high load (daytime production peak) and low load (nighttime maintenance period) to test the adaptive energy-saving ability of the system.

[0069] Table 1 records the energy consumption of different equipment during the experiment and the energy-saving effect after optimization based on the method of the present invention (both in kWh). The data compares the energy consumption of the traditional energy consumption management method and the method of the present invention.

[0070] Table 1 Experimental data table

[0071] Equipment Name Traditional Management Energy Consumption (kWh) Predicted Energy Consumption (kWh) Actual Energy Consumption after Optimization (kWh) Energy Saving Ratio (%) Load Fluctuation Rate (%) Equipment Utilization Rate (%) Central Air Conditioning System 350 330 310 11.4 5.2 87.5 Intelligent Lighting System 120 115 108 10.0 3.8 90.1 Electric Mechanical Equipment 500 470 450 10.0 6.5 85.0 Data Center Server 600 580 560 6.7 2.1 95.2 Ventilation System 200 185 180 10.0 4.3 88.3 Standby Power Supply System 150 140 138 8.0 3.5 93.0

[0072] It can be seen from the experimental data that after adopting the energy-saving prediction management system of the present invention, the actual energy consumption of each equipment has decreased significantly, and the overall energy-saving ratio has reached 9.35% on average. Compared with the traditional method, the energy consumption prediction model of the present invention can more accurately predict future load changes, enabling the energy-saving strategy to be adapted in advance and effectively reducing unnecessary energy waste. In addition, the load volatility has been reduced by more than 25% on average, indicating that the system not only optimizes energy distribution but also effectively reduces load instability and improves energy utilization efficiency.

[0073] In terms of equipment utilization, the optimization strategy of the method of the present invention has significantly improved the utilization of high-energy-consuming equipment (such as central air conditioners and data center servers). For example, the utilization rate of data center servers has reached 95.2%, which is about 3% higher than that of the traditional management mode, proving that the energy-saving strategy of the present invention not only saves energy consumption but also optimizes the operation efficiency of equipment.

[0074] In addition, this experiment also proves that the reinforcement learning optimization strategy proposed by the present invention can adapt to different load environments, automatically adjust the energy consumption management method, and make the energy-saving measures more accurate. For example, in the central air conditioning system, after optimization by reinforcement learning, the actual energy consumption is further reduced by 6% compared with the predicted energy consumption, which is difficult to achieve by traditional fixed strategies. This system dynamically optimizes the strategy based on the feedback mechanism and adjusts the reward function in combination with Bayesian optimization, so that the energy-saving strategy is continuously optimized during long-term operation, ensuring continuous energy-saving effects.

[0075] Generally speaking, the present invention has the following technical advantages compared with traditional methods:

[0076] Improve the accuracy of energy consumption prediction: The LSTA mechanism enhances the perception ability of short-term load fluctuations, reducing the prediction error.

[0077] Optimize the energy-saving strategy and reduce energy waste: The adaptive optimization of reinforcement learning makes the energy-saving strategy more intelligent, and the energy-saving ratio is increased to 9.35% on average.

[0078] Reduce load fluctuations and improve equipment utilization: The optimized energy consumption scheduling scheme reduces load instability, improves equipment utilization, and ensures the stability of energy supply.

[0079] Achieve closed-loop optimization and improve long-term energy-saving effects: Through the feedback mechanism and adaptive error analysis, the system can continuously improve the energy-saving strategy, making the energy-saving effect continuously improved during long-term operation.

[0080] This experiment fully verifies the effectiveness of the present invention in practical applications. Compared with traditional methods, it has higher energy-saving potential, more accurate energy consumption prediction ability, and more intelligent dynamic optimization mechanism, which can significantly improve the energy management level of industrial parks or intelligent buildings.

[0081] Embodiment 3, which is an embodiment of the present invention, provides an energy-saving prediction management system, including a preprocessing module 100, a prediction module 200, and an optimization module 300.

[0082] Among them, S4: The preprocessing module 100 is used to collect multi-source data of energy use terminals and perform preprocessing.

[0083] It should be noted that the preprocessing module 100 is used to collect multi-source data from the energy usage terminals and preprocess the data to ensure the accuracy of subsequent energy consumption prediction. First, this module connects to the data sources in the park and collects various data. Subsequently, through the time synchronization mechanism, the data of different devices are adjusted to a unified time reference to ensure temporal consistency. After data collection, data cleaning is performed to remove abnormal data, duplicate data, and fill in the missing data with linear interpolation to ensure data integrity. In addition, to reduce the data dimension and improve the calculation efficiency, mutual information analysis is carried out to screen out the features that have the most influence on energy consumption prediction and provide high-quality input data for the prediction module 200.

[0084] S5: The prediction module 200 is used to extract features and perform data analysis on the preprocessed data, and conduct energy consumption prediction based on a deep neural network model.

[0085] It should be noted that the prediction module 200 receives the feature data provided by the preprocessing module 100 and conducts deep learning modeling to predict the future energy consumption trend. First, this module performs time series segmentation, structuring the data according to different time windows to meet the energy consumption analysis requirements at different levels. Then, this module adopts a feature selection method based on multi-layer nested clustering to extract the global pattern and local pattern of the park's energy consumption data respectively, so as to enhance the sensitivity of the prediction model to energy consumption changes. Subsequently, based on the Laplace variational autoencoder (L-VAE), the input features are dimensionally reduced to reduce data redundancy and improve the calculation efficiency of the model. In terms of energy consumption prediction modeling, this module adopts a Transformer prediction network that combines a gated recurrent unit (GRU) and a local spatio-temporal attention mechanism (LSTA), where the GRU is used to learn the long-term energy consumption trend and the LSTA is used to enhance the modeling ability for short-term load fluctuations. Finally, this module outputs the energy consumption prediction values at different future time steps and calculates the uncertainty interval of the prediction, so that the optimization module 300 can formulate precise energy-saving strategies.

[0086] S6: The optimization module 300 is used to generate energy-saving strategies based on the prediction results and perform dynamic optimization and adjustment through a feedback mechanism.

[0087] It should be noted that the optimization module 300 formulates an energy-saving strategy based on the energy consumption prediction results generated by the prediction module 200 and performs dynamic optimization and adjustment through a feedback mechanism. First, this module establishes a multi-objective optimization model, comprehensively considering factors such as energy-saving cost, load balancing, equipment life, and user comfort, to construct an optimal energy-saving scheduling plan. In terms of strategy evaluation, this module adopts an intelligent optimization method based on reinforcement learning (DQN+MCTS), where DQN is used to continuously optimize the strategy weights, and MCTS is used to screen the globally optimal strategy. The optimized energy-saving strategies include measures such as adjusting the device operation time, intelligent load distribution, peak load reduction, and intelligent switch management to minimize the energy consumption expenditure. After the strategy is executed, this module forms a feedback data set for the execution of the energy-saving strategy by real-time monitoring the device operation data, environmental factors, and load levels. Subsequently, an adaptive error analysis model is used to calculate the deviation between the actual energy consumption value and the predicted value, and the reward function of the reinforcement learning model is adjusted in combination with Bayesian optimization, so that the optimized strategy can continuously adapt to the dynamic load changes in the park. Finally, this module redeploys the optimized energy-saving strategy to achieve the closed-loop optimization of the energy-saving strategy and ensure the long-term efficient operation of the energy management system.

[0088] If a function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this 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 causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present invention. The aforementioned storage media include: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, etc., which can store program codes.

[0089] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or used in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0090] More specific examples (a non-exhaustive list) of computer-readable media include the following: electrical connections (electronic devices) having one or more wirings, portable computer disk cartridges (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber devices, and portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which a program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing it in a suitable manner if necessary, and then storing it in a computer memory.

[0091] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, the multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A management method for energy conservation prediction, characterized in that, Including: Collecting multi-source data of energy-using terminals and performing preprocessing; Performing feature extraction and data analysis on the preprocessed data, and performing energy consumption prediction based on a deep neural network model; Generating an energy-saving strategy according to the prediction result, and performing dynamic optimization and adjustment through a feedback mechanism; The deep neural network model includes an input layer, a variational autoencoder layer, a time series modeling layer, a Transformer prediction layer, a fully connected layer, and an output layer; The input layer receives the feature set that affects the campus energy consumption prediction, and embeds it into the high-dimensional feature space through the multi-head self-attention mechanism; The variational autoencoder layer introduces a Laplace variational autoencoder to perform feature dimensionality reduction on the input campus energy consumption data, removes redundant information, optimizes data representation, regularizes the latent variables output by the encoder using the Laplace prior distribution, optimizes the variational lower bound through Laplace variational inference, combines the decoder reconstruction error and the KL divergence balance, and generates a low-dimensional interpretable feature representation; The time series modeling layer models the short-term changes of energy consumption data through a gated recurrent unit, and dynamically adjusts the hidden state in combination with the attention mechanism; The Transformer prediction layer adopts a local spatio-temporal attention mechanism, adds local time window regulation to the global Transformer architecture, dynamically divides the input sequence based on a sliding time window, applies local attention weights to the time steps within the window, adjusts the contribution ratio of global and local attention through a time weighting mechanism, and optimizes the joint modeling ability for short-term energy consumption fluctuations and long-term trends; The fully connected layer processes the temporal features output by the Transformer layer, and uses the ReLU activation function to enhance the non-linear mapping ability; The output layer converts the output of the fully connected layer into the final energy consumption prediction result; The said deep neural network model includes: The input layer receives the feature set that has the most significant impact on the campus energy consumption prediction, and uses the multi-head self-attention embedding mechanism to map the multi-dimensional input data to the high-dimensional feature space; The variational autoencoder layer introduces a Laplace variational autoencoder to perform feature dimensionality reduction on the input campus energy consumption data, removes redundant information, optimizes data representation, regularizes the latent variables output by the encoder using the Laplace prior distribution, optimizes the variational lower bound through Laplace variational inference, optimizes the modeling ability for the sparsity and mutation characteristics of energy consumption data, and combines the decoder reconstruction error and the KL divergence balance to generate a low-dimensional interpretable feature representation; The time series modeling layer uses a gated recurrent unit for time series modeling, and learns the short-term and long-term energy consumption change trends from the input temporal energy consumption data; dynamically adjusts the hidden state weights of the gated recurrent unit through the attention mechanism; The Transformer prediction layer adopts a local spatio-temporal attention mechanism, adds local time window regulation to the global Transformer architecture, dynamically divides the input sequence based on a sliding time window, applies local attention weights to the time steps within the window, and adjusts the contribution ratio of global and local attention through a time weighting mechanism, and optimizes the joint modeling ability for short-term energy consumption fluctuations and long-term trends; The fully connected layer receives the time series features output by the Transformer prediction layer and performs a non-linear transformation. The ReLU activation function is used to enhance the non-linear fitting ability; The output layer converts the output of the fully connected layer into the final energy consumption prediction result; The energy consumption prediction includes: Normalize the feature set that has the most significant impact on the campus energy consumption prediction, and convert it to the same scale using the Z-score normalization method; Use the embedding layer based on multi-head self-attention to input the feature set data of the same scale into the deep neural network to capture the long-term and short-term dependencies between different features; Based on the variational autoencoder, perform feature dimensionality reduction, convert the features into low-dimensional latent vector representations, and introduce the gated recurrent unit GRU for time series modeling to extract the temporal features of energy consumption changes; Input the output result of the GRU into the fully connected layer, and use the Softmax activation function to calculate the probability distribution of campus energy consumption at different time steps. Combine the Bayesian regression model for calculation to obtain the final campus energy consumption prediction result; The dynamic optimization adjustment through the feedback mechanism includes: Execute the optimal energy-saving strategy, continuously monitor the actual energy consumption data, obtain the real-time device operation status, environmental parameters, and load level, and form a feedback data set of the execution effect of the energy-saving strategy; Based on the feedback data set, use the adaptive error analysis model to calculate the error between the actual energy consumption value and the campus energy consumption prediction result, and analyze the error source; According to the error analysis result, dynamically optimize the energy-saving strategy.

2. The management method for energy conservation prediction according to claim 1, characterized in that: The preprocessing includes: Collect campus energy consumption data through energy usage terminals and perform time series alignment on the data based on the time synchronization mechanism; Clean the campus energy consumption data after time series alignment, including removing duplicate data, abnormal data, and filling in missing data.

3. The management method for energy conservation prediction according to claim 1 or 2, characterized in that: The feature extraction and data analysis include: Segment the preprocessed campus energy consumption data into time series. Use the feature selection method based on multi-layer nested clustering to classify and summarize the global and local patterns of the campus energy consumption data respectively, and combine the mutual information analysis to calculate the mutual influence degree between feature variables, and screen the feature set that has the most significant impact on the campus energy consumption prediction.

4. The management method for energy conservation prediction according to claim 3, characterized in that: Generating an energy-saving strategy based on the prediction result includes: Calculate and generate a campus energy consumption energy-saving strategy according to the campus energy consumption prediction result. Use the reinforcement learning algorithm to evaluate the adaptability of different strategies and screen the optimal energy-saving strategy.

5. A system adopting the management method of energy-saving prediction as described in any one of claims 1 to 4, characterized in that: It includes a preprocessing module (100), a prediction module (200), and an optimization module (300); The preprocessing module (100) is used to collect multi-source data from energy usage terminals and perform preprocessing; The prediction module (200) is used to perform feature extraction and data analysis on the preprocessed data, and perform energy consumption prediction based on the deep neural network model; The optimization module (300) is used to generate an energy-saving strategy based on the prediction result and perform dynamic optimization adjustment through the feedback mechanism.

6. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the energy-saving prediction management method described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the energy-saving prediction management method described in any one of claims 1 to 4.

Images