Household load prediction method based on empirical mode decomposition and Bayesian optimization
Through methods based on empirical modal decomposition and Bayesian optimization, historical load data are subdivided and expanded, and convolutional neural network and long and short-term memory unit model are constructed, the problem of load prediction error and interdependence in the existing technology is solved, and more accurate energy demand prediction is achieved.
Patent Information
- Application Number
- CN202411832048.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2025-04-15
AI Technical Summary
The prior art has accumulated errors in residential load prediction and neglected interdependence between input and output variables, resulting in inaccurate prediction results.
The family load prediction method based on empirical modal decomposition and Bayesian optimization is adopted to subdivide historical load data through empirical modal decomposition, expand feature data, and build a model using convolutional neural network and long and short-term memory units. The hyperparameters are optimized through Bayesian optimization method, and the deviation correction is finally performed.
Multi-output prediction of energy demand is achieved, the correlation between each input variable is comprehensively taken into account, and the prediction effect is significantly improved, surpassing traditional direct or recursive strategies.
Smart Images
Figure CN119834209A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of load forecasting, and in particular to a household load forecasting method based on empirical mode decomposition and Bayesian optimization. Background Art
[0002] As renewable energy accounts for a growing share of energy systems, accurate forecasting of residential loads has become an urgent need to ensure effective management, planning strategies, and policy formulation. Given the highly random nature of energy load characteristics, accurate forecasting faces many challenges, and forecast results often contain errors.
[0003] When using traditional multi-step forecasting strategies, including direct and recursive methods, forecasting models are often limited by error accumulation and neglect of the interdependence between input and output variables. It is worth noting that existing related patents mostly rely on directly modeling the original demand profile, which ignores the intrinsic characteristics of the data to a certain extent. In addition, in order to capture the complex patterns in the data set and generate accurate forecasts, it is necessary to build a complex model architecture, which undoubtedly increases the difficulty of model convergence and may lead to an increase in forecast errors. Therefore, it is necessary to provide a household load forecasting method based on empirical mode decomposition and Bayesian optimization to solve the above problems. Summary of the invention
[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a household load forecasting method based on empirical mode decomposition and Bayesian optimization, with the goal of improving the prediction performance of the machine learning model by systematically applying the currently popular and proven machine learning technology.
[0005] The present invention is implemented as follows: a household load forecasting method based on empirical mode decomposition and Bayesian optimization, the method comprising the following steps:
[0006] Collecting historical load data and performing data preprocessing on the historical load data, wherein the data preprocessing includes data filling, outlier processing and logarithmic transformation;
[0007] Performing empirical mode decomposition to subdivide the historical load data into N intrinsic mode functions, where N is a constant value;
[0008] Perform feature engineering to augment feature data for historical load data decomposed into N intrinsic mode functions;
[0009] The model is built based on convolutional neural networks and long short-term memory units, using a multi-input architecture. There are two convolutional layers in the model structure to extract features from lagged observation data. A maximum pooling layer is added after each convolutional layer to avoid overfitting.
[0010] The hyperparameters of the constructed model are optimized by Bayesian optimization method;
[0011] After completing the model prediction and obtaining N decomposed prediction values, a reverse decomposition operation is performed to merge the prediction values, and the bias correction is performed on the merged prediction values.
[0012] As a further solution of the present invention: the step of preprocessing the historical load data specifically includes:
[0013] For missing data, forward filling is used to supplement it. When the number of consecutive missing data points exceeds three, the average value of the corresponding points four days before the missing point is collected to fill the gap.
[0014] Outliers were identified by the interquartile range method, and once identified, they were processed using quantile-based lower and upper limits;
[0015] Logarithmic transformation was performed to ensure normality of the data and reduce data variability.
[0016] As a further solution of the present invention: the step of expanding the characteristic data for the historical load data decomposed into N intrinsic mode functions specifically includes:
[0017] Add lagging observation features to historical load data, and add the previous ten points of historical data as features for each point of data;
[0018] Time series features are introduced to enhance the prediction ability of the model. Time series features include month, hour, minute and time period classification. Time period classification adopts one-hot encoding.
[0019] Calculate statistical features based on previous observations, the statistical features including mean, standard deviation, minimum, maximum, median, interquartile range, skewness, kurtosis, energy of the sequence, and the number of peaks in the sequence;
[0020] Add Fourier transform features. For each time step, perform Fourier transform on the previous observations and calculate the Fourier transformed features, including mean, standard deviation, minimum, maximum, median, interquartile range, skewness, kurtosis, energy of the sequence, and the number of peaks in the sequence.
[0021] As a further solution of the present invention: when building the model, after the output of the convolutional layer is flattened, the information is passed to two bidirectional LSTM layers to detect patterns in two directions; the bidirectional LSTM is a subset of the LSTM unit, consisting of two independent units, and within each time step, the bidirectional structure allows the network to have both backward and forward knowledge about the sequence.
[0022] As a further solution of the present invention: when building the model, after the information passes through the random inactivation layer, the first input array reaches a node and is combined with the engineering features. The engineering features will pass through three dense layers and reach the node; the output of the node will enter two fully connected dense layers for prediction; in the pooling stage of the network, maximum pooling is selected with a pool size of 2, the first three dense layers have 128 neurons each, and the last two dense layers are responsible for mapping the input data to the predicted values, with 50 and 25 neurons respectively.
[0023] As a further solution of the present invention: when the constructed model is optimized for hyperparameters by the Bayesian optimization method, the specific parameters are set as follows: the learning rate adjustment range is limited to between 10e-5 and 10e-1; the L2 regularization parameter is set between 0.1 and 1; the random inactivation rate ranges from 0.3 to 0.7; the batch size ranges from 32 to 1024 and must be a multiple of 32; the number of nodes in the LSTM network is set between 32 and 256.
[0024] As a further solution of the present invention: the step of performing bias correction on the merged predicted values specifically includes: calculating all percentiles of the observation series and the prediction series from the 0th percentile to the 100th percentile with a unit step size; calculating the difference; and performing bias correction on each predicted value according to the percentile range using a percentile-based bias correction difference item.
[0025] Compared with the prior art, the present invention has the following beneficial effects:
[0026] The present invention performs empirical mode decomposition to subdivide the historical load data into N intrinsic mode functions, and expands the characteristic data of the historical load data decomposed into N intrinsic mode functions; then constructs a model based on a convolutional neural network and a long short-term memory unit; optimizes the hyperparameters of the constructed model by the Bayesian optimization method; after completing the model prediction and obtaining the N decomposed prediction values, performs an inverse decomposition operation to merge the prediction values, and performs deviation correction on the merged prediction values. The multi-output prediction of energy demand is realized, which not only fully considers the correlation between the input variables, but also significantly surpasses the traditional direct or recursive strategies in terms of prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 Flowchart of a household load forecasting method based on empirical mode decomposition and Bayesian optimization.
[0028] Figure 2 This is a model architecture diagram for a household load forecasting method based on empirical mode decomposition and Bayesian optimization. DETAILED DESCRIPTION
[0029] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0030] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.
[0031] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a household load forecasting method based on empirical mode decomposition and Bayesian optimization, and the method includes the following steps:
[0032] Step 1: Collect historical load data and perform data preprocessing on the historical load data, wherein the data preprocessing includes data filling, outlier processing and logarithmic transformation;
[0033] Step 2: perform empirical mode decomposition to subdivide the historical load data into N intrinsic mode functions (IMFs), where N is a constant value, and here N is 6, that is, subdivided into six IMFs;
[0034] Step 3: Perform feature engineering to expand feature data for the historical load data decomposed into N intrinsic mode functions;
[0035] Step 4: Build a model based on convolutional neural networks and long short-term memory units, using a multi-input architecture. There are two convolutional layers in the model structure to extract features from lagged observation data. A maximum pooling layer is added after each convolutional layer to avoid overfitting; and reduce computational costs by limiting the number of parameters that need to be learned. Maximum pooling is a sample-based discretization technique that aims to minimize the dimension of the input vector. The program uses the pooling layer in the architecture to return the features with the highest value and importance.
[0036] Step 5: Optimize the hyperparameters of the constructed model through Bayesian optimization method;
[0037] Step six, perform bias correction. After completing the model prediction and obtaining N decomposed prediction values, perform an inverse decomposition operation to merge the prediction values, and perform bias correction on the merged prediction values.
[0038] In the embodiment of the present invention, a deep neural network is selected, which is specifically implemented as a combination of a convolutional neural network and a long short-term memory unit (CNN-LSTM) to perform multi-output prediction of energy demand. This model not only comprehensively considers the correlation between the input variables, but also significantly surpasses the traditional direct or recursive strategies in terms of prediction effect. A hybrid modeling framework is also adopted. The framework includes a comprehensive decomposition module, namely Empirical Mode Decomposition (EMD), for data preprocessing, and an independent module that specifically performs a deep learning process in order to provide a new solution.
[0039] In the embodiment of the present invention, the step of preprocessing the historical load data specifically includes: supplementing the missing data by forward filling, when the number of consecutive missing data points exceeds three, the average value of the corresponding points four days before the missing point will be collected for filling; after completing the missing value filling, outliers are identified by the interquartile range (IQR) method, IQR is an indicator to measure statistical dispersion, and the specific calculation method is the 75th percentile minus the 25th percentile. Within the IQR range, if the value falls between -1.5 and 1.5 times the IQR, it is usually considered normal. In the embodiment of the present invention, 1.5 is selected as this factor. After the outliers are identified, the lower and upper limits based on the quantiles are used for processing; specifically, the lower limit is set by the 2nd percentile, and the upper limit is set by the 98th percentile. The purpose of applying this technology is not only to eliminate outliers, but also to maintain the integrity of the core demand model of the basic data. The selected standard aims to achieve a balance between modifying the data and maintaining its integrity. Finally, logarithmic transformation is performed to ensure the normality of the data and reduce the variability of the data.
[0040] In an embodiment of the present invention, the process of performing empirical mode decomposition (EMD) is based on a core assumption that the time series is composed of multiple simple intrinsic oscillation modes, which are defined as intrinsic mode functions (IMFs). Each IMF is closely related to a specific frequency range derived from the original data. In short, these IMFs constitute amplitude modulation-frequency modulation (AM-FM) signals, which reflect different frequency bands in the original time series, gradually transitioning from high frequency (first IMF) to low frequency (last IMF). In this solution, this algorithm will be used to subdivide the load history data into six IMFs (using the pyEDM tool to directly perform the decomposition) to ensure that the amount of data required for model training is fully considered, while ensuring that the historical load data of each household can be clearly and accurately decomposed.
[0041] In the embodiment of the present invention, the step of expanding the feature data of the historical load data decomposed into N intrinsic mode functions specifically includes: adding lagged observation features to the historical load data, adding the historical data of the first ten points as features for the data of each point; introducing time series characteristics to enhance the prediction ability of the model, the time series characteristics include but are not limited to the time period classification of months, hours, minutes and time points, the time period classification adopts unique hot encoding, and the specific time periods are divided into dawn (4, 5), early morning (6, 7), morning (8, 9, 10), noon (11, 12, 13), afternoon (14, 15, 16), evening (17, 18, 19), night (20, 21, 22) and midnight (23, 0, 1, 2, 3), in addition, factors such as the day of the week, whether it is a holiday, whether it is a weekend, etc. will also be considered. Finally, in order to improve the accuracy of the model, the cosine and sine values of the hour will be introduced as additional features. Then, statistical features are calculated based on the 96 previous observations in the data, aiming to optimize the performance of the model so that it can more accurately capture the potential patterns of the previous two days. The statistical features include mean, standard deviation, minimum, maximum, median, interquartile range, skewness, kurtosis, energy of the sequence, and the number of peaks in the sequence. In addition, in order to enrich the feature set, moving averages with windows of 10, 20, 30, and 40, and statistics such as the mean, minimum, and maximum of the expanded window are added. Finally, Fourier transform features are added to the data. Fourier transform provides a different perspective for analyzing time signals because some signal properties can be fully explored in the frequency domain. For each time step, the previous 96 observations are Fourier transformed, and the Fourier transformed features are calculated, including mean, standard deviation, minimum, maximum, median, interquartile range, skewness, kurtosis, energy of the sequence, and the number of peaks in the sequence.
[0042] In the embodiment of the present invention, when constructing the model, after the output of the convolutional layer is flattened, the information is passed to two bidirectional LSTM layers to detect patterns in two directions; the bidirectional LSTM is a subset of the LSTM unit, consisting of two independent units, and within each time step, the bidirectional structure allows the network to have both backward and forward knowledge about the sequence. After the information passes through the dropout layer, the first input array arrives at a node and is combined with the engineering features. The engineering features will pass through three dense layers and reach the node; the output of the node will enter two fully connected dense layers for prediction, and the activation function of these layers is the default recommended linear rectifier unit (ReLU) function; in the pooling stage of the network, maximum pooling is selected with a pool size of 2. The first three dense layers have 128 neurons each, and the last two dense layers are responsible for mapping the input data to the predicted values, with 50 and 25 neurons respectively. These layers use the ReLU function as the activation function. In addition to the dropout layer used in the network, L 2Regularization. 2 The parameter is usually called weight decay, which brings the weights closer to the origin by adding a regularization term to the objective function. When compiling the model, the Adaptive Moment Estimation (Adam) algorithm is used as the gradient optimizer for the architecture.
[0043] In an embodiment of the present invention, when the constructed model is optimized for hyperparameters by the Bayesian optimization method, the specific parameters are set as follows: the learning rate adjustment range is limited to between 10e-5 and 10e-1; the L2 regularization parameter is set between 0.1 and 1; the random inactivation rate ranges from 0.3 to 0.7; the batch size value is between 32 and 1024, and must be a multiple of 32; the number of nodes of the LSTM network is set between 32 and 256. The above parameter settings are intended to ensure the stability and optimization effect of the model.
[0044] In an embodiment of the present invention, the step of performing bias correction on the merged predicted values specifically includes: calculating all percentiles of the observation series and the prediction series from the 0th percentile to the 100th percentile with a unit step size; calculating the difference; and performing bias correction on each predicted value according to the percentile range using a percentile-based bias correction difference item.
[0045] In summary, a unique framework or process integrating multiple mature data science methods has been constructed. The construction concept of this new framework stems from the improvement of the integration capabilities of various methods, thus covering widely used data science methods, suitable for systematic processing of data preprocessing, modeling and post-processing tasks. The use of the framework is innovative because it uniquely combines multiple data science methods. Specifically, the framework aims to verify the effectiveness of empirical mode decomposition (EMD) as a preprocessing technique in demand decomposition, and combines it with long short-term memory networks (LSTM) and convolutional neural networks (CNN), with strategic post-processing solutions to improve the accuracy of demand forecasting.
[0046] The above only describes in detail the preferred embodiments of the present invention, which is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
[0047] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.
[0048] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
[0049] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the disclosure in the specification and examples. This application is intended to cover any variations, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.
Claims
1. A household load forecasting method based on empirical mode decomposition and Bayesian optimization, characterized in that: The method comprises the following steps: Collecting historical load data and performing data preprocessing on the historical load data, wherein the data preprocessing includes data filling, outlier processing and logarithmic transformation; Performing empirical mode decomposition to subdivide the historical load data into N intrinsic mode functions, where N is a constant value; Perform feature engineering to augment feature data for historical load data decomposed into N intrinsic mode functions; The model is built based on convolutional neural networks and long short-term memory units, using a multi-input architecture. There are two convolutional layers in the model structure to extract features from lagged observation data. A maximum pooling layer is added after each convolutional layer to avoid overfitting. The hyperparameters of the constructed model are optimized by Bayesian optimization method; After completing the model prediction and obtaining N decomposed prediction values, a reverse decomposition operation is performed to merge the prediction values, and the bias correction is performed on the merged prediction values.
2. The household load forecasting method based on empirical mode decomposition and Bayesian optimization according to claim 1, characterized in that: The step of preprocessing the historical load data specifically includes: For missing data, forward filling is used to supplement it. When the number of consecutive missing data points exceeds three, the average value of the corresponding points four days before the missing point is collected to fill the gap. Outliers were identified by the interquartile range method, and once identified, they were processed using quantile-based lower and upper limits; Logarithmic transformation was performed to ensure normality of the data and reduce data variability.
3. The household load forecasting method based on empirical mode decomposition and Bayesian optimization according to claim 1, characterized in that: The step of expanding characteristic data for the historical load data decomposed into N intrinsic mode functions specifically includes: Add lagging observation features to historical load data, and add the previous ten points of historical data as features for each point of data; Time series features are introduced to enhance the prediction ability of the model. Time series features include month, hour, minute and time period classification. Time period classification adopts one-hot encoding. Calculate statistical features based on previous observations, the statistical features including mean, standard deviation, minimum, maximum, median, interquartile range, skewness, kurtosis, energy of the sequence, and the number of peaks in the sequence; Add Fourier transform features. For each time step, perform Fourier transform on the previous observations and calculate the Fourier transformed features, including mean, standard deviation, minimum, maximum, median, interquartile range, skewness, kurtosis, energy of the sequence, and the number of peaks in the sequence.
4. The household load forecasting method based on empirical mode decomposition and Bayesian optimization according to claim 1, characterized in that: When building the model, after the output of the convolutional layer is flattened, the information is passed to two bidirectional LSTM layers to detect patterns in both directions; bidirectional LSTM is a subset of LSTM units that consists of two independent units. Within each time step, the bidirectional structure allows the network to have both backward and forward knowledge about the sequence.
5. The household load forecasting method based on empirical mode decomposition and Bayesian optimization according to claim 4, characterized in that: When building the model, after the information passes through the random dropout layer, the first input array reaches a node and is combined with the engineering features, which will pass through three dense layers to reach the node; the output of the node will enter two fully connected dense layers for prediction; In the pooling stage of the network, max pooling was selected with a pool size of 2. The first three dense layers had 128 neurons each, and the last two dense layers were responsible for mapping the input data to the predicted values, with 50 and 25 neurons respectively.
6. The household load forecasting method based on empirical mode decomposition and Bayesian optimization according to claim 1, characterized in that: When the hyperparameters of the constructed model are optimized by the Bayesian optimization method, the specific parameter settings are as follows: the learning rate adjustment range is limited to 10e-5 to 10e-1; the L2 regularization parameter is set between 0.1 and 1; the random inactivation rate ranges from 0.3 to 0.7; the batch size ranges from 32 to 1024 and must be a multiple of 32; the number of nodes in the LSTM network is set between 32 and 256.
7. The household load forecasting method based on empirical mode decomposition and Bayesian optimization according to claim 1, characterized in that: The step of performing bias correction on the combined predicted values specifically includes: calculating all percentiles of the observed series and the predicted series from the 0th percentile to the 100th percentile with a unit step size; calculating the difference; and performing bias correction on each predicted value according to the percentile range using a percentile-based bias correction difference term.