Power load prediction method

Abstract

The invention is suitable for the technical field of power load prediction, and provides a power load prediction method. The method comprises the following steps: acquiring original power load and original meteorological data; the original power loads are in one-to-one correspondence with the original meteorological data; performing adaptive decomposition on the original power load based on the original meteorological data to obtain a meteorological load and a basic load; the meteorological load represents a random fluctuation load driven by meteorological factors; predicting according to the meteorological load and the original meteorological data by using a pre-constructed TimeMixer model to obtain a meteorological load prediction result; and carrying out superposition reconstruction on the meteorological load prediction result and the basic load to obtain a day-ahead power load prediction result. According to the power load prediction method, full mining of the weather-load association relationship is realized, the problem of complex multi-source association between the weather factor and the power load is solved, and the prediction precision is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power load prediction, and particularly relates to a power load prediction method. BACKGROUND

[0002] With the intensification of global climate change, extreme high temperature, cold wave and other abnormal weather events occur frequently, resulting in significant increase in the volatility and uncertainty of power system load due to meteorological factors. Under extreme temperature conditions, air conditioning and heating load rises sharply, which easily leads to expansion of power grid peak-valley difference, local equipment overload and even power supply gap. At the same time, the output of wind power, photovoltaic and other new energy sources is also affected by meteorological mutations and fluctuates sharply, further increasing the difficulty of power grid balance and dispatching operation, and posing a serious threat to the safe and economic operation of the power system.

[0003] To cope with the above challenges, short-term meteorological load prediction has become one of the key technologies for power system operation planning. Existing methods usually based on temperature, humidity, wind speed and solar radiation intensity and other basic meteorological elements, through statistical model or machine learning algorithm, build the mapping relationship between them and historical load data, so as to realize the prediction of future load change. To some extent, this kind of technology can reflect the influence of meteorological conditions on conventional load, and is an important basis for current generation plan formulation and dispatching decision.

[0004] However, the existing short-term load prediction method has not fully solved the complex multi-source correlation problem between meteorological factors and power load when coping with extreme weather and high proportion of new energy access scenarios, and the prediction accuracy is limited. Therefore, it is urgent to develop a more accurate short-term meteorological load prediction method to effectively support power grid safety warning, generation plan scheduling optimization and power market decision. SUMMARY

[0005] To overcome the problems in the related art, the embodiments of the application provide a power load prediction method, which can solve the complex multi-source correlation problem between meteorological factors and power load and improve the prediction accuracy.

[0006] The application is realized by the following technical solutions: The embodiments of the application provide a power load prediction method, which comprises: Obtaining original power load and original meteorological data; each original power load corresponds to each original meteorological data; Based on the original meteorological data, the original power load is adaptively decomposed to obtain meteorological load and basic load; the meteorological load represents random fluctuation load driven by meteorological factors; Using a pre-constructed TimeMixer model, meteorological load prediction results are predicted according to the meteorological load and the original meteorological data; Superimpose the meteorological load prediction result and the basic load to reconstruct a day-ahead power load prediction result.

[0007] In a possible implementation, the original power load is adaptively decomposed based on original meteorological data to obtain the meteorological load and the basic load, including: The original power load is preprocessed to obtain a preprocessed power load; The preprocessed power load is adaptively decomposed to obtain a plurality of component sequences of different frequency bands; An adaptive load type criterion is determined based on the original meteorological data and the plurality of component sequences of different frequency bands; The meteorological load and the basic load are determined based on the adaptive load type criterion and the plurality of component sequences of different frequency bands.

[0008] In a possible implementation, the adaptive load type criterion is determined based on the original meteorological data and the plurality of component sequences of different frequency bands, including: Correlation coefficients of the component sequences of different frequency bands and meteorological variables in the original meteorological data are calculated; A correlation coefficient change curve is drawn based on the correlation coefficients; An inflection point at which the correlation coefficient appears to decline in the correlation coefficient change curve is determined; and the inflection point is taken as the adaptive load type criterion.

[0009] In a possible implementation, the meteorological load and the basic load are determined based on the adaptive load type criterion and the plurality of component sequences of different frequency bands, including: The component sequences of the m frequency bands before the adaptive load type criterion are reconstructed as the meteorological load, and the component sequences of the n frequency bands after the adaptive load type criterion and a residual term are reconstructed as the basic load.

[0010] In a possible implementation, a TimeMixer model is constructed in advance, and a meteorological load prediction result is predicted based on the meteorological load and the original meteorological data, including: For any one meteorological feature, a Pearson correlation coefficient between the meteorological feature and the meteorological load is calculated; wherein the original meteorological data includes a plurality of meteorological features; Strongly correlated meteorological features are screened based on the Pearson correlation coefficients; The meteorological load and the strongly correlated meteorological features are input into the TimeMixer model constructed in advance for targeted prediction, and a meteorological load prediction result is output.

[0011] In a possible implementation, the preprocessed power load is adaptively decomposed to obtain a plurality of component sequences of different frequency bands, including: The CEEMDAN algorithm is adopted to adaptively decompose the preprocessed power load, to obtain component sequences of multiple different frequency bands.

[0012] In a possible implementation, the CEEMDAN algorithm is adopted to adaptively decompose the preprocessed power load, to obtain component sequences of multiple different frequency bands, including: The original power load sequence is determined based on the preprocessed power load. Gaussian white noise of different amplitudes is superimposed in the original power load sequence respectively, to construct multiple noise auxiliary sequences. The CEEMD is decomposed on each group of noise auxiliary sequences respectively, to obtain decomposition results of each time; the decomposition results of each time include corresponding IMF components. The corresponding IMF components are integrated and averaged, to obtain an IMF component set.

[0013] In a possible implementation, the TimeMixer model includes a multi-scale sequence generation module, a past-decomposable-mixing (PDM) module, and a future-multipredictor-mixing (FMM) module. The multi-scale sequence generation module is configured to perform multi-scale division on the input meteorological load and original meteorological data, to generate scale sequences respectively representing short-term, medium-term, and long-term dynamics. The PDM module is configured to perform feature extraction on the scale sequences respectively, to obtain scale information subjected to feature extraction. The FMM module is configured to integrate the scale information subjected to feature extraction, and output a meteorological load prediction result.

[0014] In a possible implementation, the component sequences of the first m frequency bands before the adaptive load type criterion are high-frequency IMF components; the component sequences of the last n frequency bands and the residual term after the adaptive load type criterion are low-frequency IMF components.

[0015] In a possible implementation, the preprocessing of the original power load includes: The original power load is processed by using the Z-score standardization; The missing data of the original power load is filled by using the linear interpolation method.

[0016] Compared with the prior art, the embodiments of the present application have the beneficial effects that: The embodiments of the present application decompose and separate the load component purely driven by meteorology through self-adaptation, and deeply integrate multi-source meteorological data by using a TimeMixer model, specially model the complex correlation therebetween, and realize full mining of the meteorology-load correlation. By decomposing the complex comprehensive prediction problem into two sub-problems of relatively stable basic load prediction and highly dynamic meteorological load prediction, and respectively processing them by using optimal strategies, error superposition and confusion are eliminated, the prediction accuracy is fundamentally improved, the prediction robustness and accuracy under a high proportion of new energy and an extreme weather scenario are significantly improved, the prediction result can more effectively serve advanced applications such as power grid safety early warning, power generation plan scheduling optimization and power market decision, and has great engineering practical value.

[0017] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the present specification. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 is a flowchart of a power load prediction method provided by an embodiment of the present application; Figure 2 is a flowchart of step 102 provided by an embodiment of the present application; Figure 3 is a flowchart of step 103 provided by an embodiment of the present application; Figure 4 is a functional flowchart of the overall implementation of the TimeMixer model provided by an embodiment of the present application; Figure 5 is a time mixer structure provided by an embodiment of the present application; Figure 6 is a basic load, meteorological load and residual item curve diagram provided by an embodiment of the present application; Figure 7 is a correlation heat map of meteorological load and each meteorological variable provided by an embodiment of the present application; Figure 8 is a prediction curve of each model under the same condition provided by an embodiment of the present application; Figure 9 is a structural schematic diagram of a controller provided by an embodiment of the present application. DETAILED DESCRIPTION

[0020] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known methods, devices, circuits, and

[0021] It is to be understood that the terminology "includes", "has", "holds", "contains" and / or "comprising", "including", "having" and their conjugates, as used herein, means "including but not limited to", and not to the exclusion of any other term or aspect.

[0022] It is also to be understood that the terminology "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items, as well as the simple inclusion of the items.

[0023] As used in the description of the application and the appended claims, the term "if' can be interpreted to mean "when" or "upon" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if it is determined" or "if [a described condition or event] is detected" can be interpreted to mean "upon determining" or "in response to determining" or "upon detecting [the described condition or event]" or "in response to detecting [the described condition or event]".

[0024] In addition, the terms "first", "second", "third", etc. as used in the description and the appended claims are used only to distinguish one element from another and do not otherwise limit the scope of the invention.

[0025] Reference throughout this specification to "an embodiment" or "some embodiments" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. Thus, the appearances of the phrases "in one embodiment", "in some embodiments", "in other embodiments", "in additional embodiments", and so on, in various places throughout this specification are not necessarily all referring to the same embodiment, unless otherwise specified. The terms "comprise", "comprising", "has", "having", "includes" and "including" and conjugations thereof, are used throughout this specification, and are by definition open- ended, meaning that they include the stated feature, element, step, component, or function, but do not exclude others.

[0026] In order to better understand the present application, a further introduction to the present application is provided.

[0027] The current weather load prediction methods cover a variety of types from traditional time series models (such as ARIMA, SARIMA) to machine learning and deep learning methods (such as support vector regression, long short-term memory network LSTM, convolutional neural network CNN) and the like. Traditional statistical models perform well in processing linear features and short-term stationary sequences, but it is difficult to capture nonlinear and multi-scale dynamic characteristics. Although deep learning models have strong nonlinear modeling capabilities, they are sensitive to mixed noise and feature redundancy in multi-source data, which affects the prediction accuracy.

[0028] To solve these problems, signal decomposition techniques (such as variational mode decomposition VMD) are used to reduce the complexity of the load signal, and feature selection and advanced prediction models are combined to improve prediction performance. However, existing research has not fully addressed the complex multi-source correlation between weather factors and power load, and the current prediction model still needs to be optimized for multi-scale component information.

[0029] Currently, the inventors have found that the prediction method mainly faces two major bottlenecks: first, traditional models usually directly couple weather data with original load data, and fail to effectively distinguish the sensitive part of the load affected by weather factors from the rigid basic part determined by social macro demand, resulting in mixed learning objectives and limited prediction accuracy; second, even if signal decomposition techniques are used, they are mostly treated as noise reduction means, and the decomposed components are still treated as ordinary time series signals, without establishing a clear and targeted physical correlation with weather factors, making the selection of prediction models lack sufficient basis and targeting.

[0030] The core technical problem to be solved by the present application is how to accurately isolate and quantify the part of the load fluctuation caused purely by weather changes (i.e. "weather load" in the present application) from the mixed power load signal, and to construct a dedicated and high-precision prediction model for the unique dynamic characteristics of this part of the load, thereby fundamentally improving the accuracy and reliability of the day-ahead weather load prediction. The present application proposes an innovative short-term weather load prediction method - a load prediction method based on CEEMDAN decomposition and TimeMixer model, the core logic of which is to gradually improve the prediction accuracy of weather load through the process of "physical decomposition - targeted prediction - accurate reconstruction", thereby improving the prediction accuracy of the final day-ahead power load. The comparison experiments of the TimeMixer model in the present application with a variety of benchmark models show that the method of the present application can more effectively capture the dynamic influence of weather conditions on load changes, and is superior to traditional models in multiple evaluation indicators, providing reliable technical support for power system dispatching and energy management.

[0031] The present application will be further described in detail below in conjunction with the drawings and specific embodiments.

[0032] Figure 1is a flowchart of an electric load prediction method provided by an embodiment of the present application, referring to Figure 1 The detailed description of the electric load prediction method is as follows: The embodiment of the present application provides an electric load prediction method, which comprises the following steps: Step 101, obtaining original electric load and original meteorological data.

[0033] The original electric load data is usually obtained from the energy management system, the advanced measurement system or the smart meter of the power grid dispatching center. The original electric load data is the total active power value of one or more regions in the power grid, which is recorded at a fixed time interval to form a load time series. Typical time resolution includes 15 minutes, 30 minutes or 1 hour. Each data point should contain a time stamp and a corresponding load value.

[0034] The original meteorological data can come from a meteorological observation station, a reanalysis database (such as ERA5) or a numerical weather prediction (NWP) model. The original meteorological data contains various meteorological elements that have a strong correlation with the electric load, at least including temperature, humidity, light intensity, wind speed, etc., and other elements such as weather type (sunny, rainy, snowy, etc.), cloud cover, air pressure, etc.

[0035] Each original electric load corresponds to each original meteorological data, that is, the electric load data and the meteorological data are completely synchronized in the time dimension, providing a reliable data basis for subsequent analysis.

[0036] Step 102, based on the original meteorological data, the original electric load is adaptively decomposed to obtain a meteorological load and a basic load.

[0037] The meteorological load represents the random fluctuation load driven by meteorological factors.

[0038] For example, by preprocessing the original meteorological data, the CEEMDAN algorithm is used to adaptively decompose the preprocessed electric load sequence to obtain a plurality of intrinsic mode functions (IMF) of different frequencies. According to the physical meaning, the components are reconstructed: the high-frequency IMF is defined as the "meteorological load", which represents the random fluctuation driven by meteorological factors; the low-frequency IMF and the residual term are reconstructed as the "basic load", which represents the stable basic load formed by the law of social production and life, thereby realizing the physical decoupling of the complex load signal and laying a foundation for subsequent targeted modeling.

[0039] Step 103, using a pre-constructed TimeMixer model, a meteorological load prediction result is predicted according to the meteorological load and the original meteorological data.

[0040] For example, for meteorological load, meteorological correlation analysis and feature screening are carried out. By calculating the Pearson correlation coefficient of meteorological load and temperature, humidity, irradiance, wind speed and other meteorological variables, and supplemented by heat map visualization, key meteorological features are selected as model inputs to reduce redundancy and improve prediction efficiency.

[0041] Before prediction, the TimeMixer model is constructed and trained for targeted prediction. Then the screened meteorological features and meteorological load are input into the model to output accurate prediction of meteorological load. In the whole prediction process, CEEMDAN decomposition provides pure target input for TimeMixer, and TimeMixer realizes accurate fitting of meteorological load by virtue of its multi-scale mixing mechanism. The two are deeply coordinated to form an end-to-end optimized prediction link.

[0042] In step 104, the meteorological load prediction result is superimposed and reconstructed with the basic load to obtain the day-ahead power load prediction result.

[0043] For example, the predicted meteorological load is superimposed and reconstructed with the basic load to obtain a high-precision day-ahead total power load prediction value, thereby providing reliable support for power system dispatching decision-making.

[0044] For example, ensure that the meteorological load prediction sequence and the basic load sequence are completely aligned in the time dimension, perform algebraic addition of the prediction values at the corresponding time, obtain the total load prediction value at the time, and combine the addition results of all time to obtain the final day-ahead power load prediction result. A lightweight machine learning model can also be used to learn the superposition error and correct it to obtain the total load prediction value at different times. It should be noted that the superposition reconstruction method is not limited to the above method, and a suitable superposition reconstruction method can be selected based on different scene accuracy requirements.

[0045] In an embodiment of the present application, the original power load sequence is first adaptively decomposed to decompose the complex load signal into basic load and meteorological load, thereby effectively removing the interference of non-meteorological factors. Subsequently, through correlation coefficient analysis, meteorological features (such as temperature, irradiance, cloud cover, etc.) that are significantly correlated with meteorological load are selected as model inputs. Finally, a multi-scale prediction model TimeMixer designed for time series is used to predict meteorological load with high precision. This model can capture both micro fluctuations and macro trends in the sequence, and further improve prediction performance through multi-scale information fusion. The process logic of steps 101-104 is clear, and the technologies in each link are closely coordinated, taking into account physical interpretability and prediction accuracy.

[0046] In an embodiment, Figure 2 FIG. 2 is a flowchart of step 102 according to an embodiment of the present application, which is described with reference to Figure 2The detailed process of adaptively decomposing the original power load based on the original meteorological data to obtain the meteorological load and the basic load will be described in detail.

[0047] Then, step 102 includes: Step 1021, preprocessing the original power load to obtain a preprocessed power load.

[0048] For example, preprocessing the original power load includes: using Z-score standardization to process the original power load; using linear interpolation to fill in the missing data of the original power load.

[0049] When the original power load is processed by the Z-score standardization method, the Z-score value of each data point of the original power load is calculated. The threshold range of the Z-score value can be set as [-3, +3], and the data points exceeding the range are determined as abnormal values and removed from the data set. The missing data of the original power load is generally caused by sensor failure or communication interruption, which is ingeniously repaired by linear interpolation to fill in the data sequence caused by hardware failure, ensuring the continuity of the load curve in the time dimension.

[0050] These preprocessing operations work together to convert the original power load data into a "standardized" data set, ensuring the quality and consistency of the input data.

[0051] Step 1022, adaptively decomposing the preprocessed power load to obtain a plurality of component sequences of different frequency bands.

[0052] Step 1022 includes: Adaptively decomposing the preprocessed power load by the CEEMDAN algorithm to obtain a plurality of component sequences of different frequency bands.

[0053] For example, adaptively decomposing the preprocessed power load by the CEEMDAN algorithm to obtain a plurality of component sequences of different frequency bands includes: Based on the preprocessed power load, determine the original power load sequence; superimpose different amplitude Gaussian white noise in the original power load sequence to construct a plurality of noise auxiliary sequences; then, perform CEEMD decomposition on each group of noise auxiliary sequences to obtain each decomposition result; wherein each decomposition result includes corresponding IMF components; integrate and average the corresponding IMF components to obtain an IMF component set.

[0054] Specifically, assuming that the original power load sequence is , the Gaussian white noise is , the noise amplitude coefficient is , the first The decomposition yields the th... The order of IMF is denoted as IMF_ EMD extraction of the first The operator of the order IMF is denoted as .

[0055] Noise-assisted sequences were obtained by repeatedly adding independent white noise to the signal and performing EMD analysis. And perform first-order IMF extraction to obtain the first-order IMF:

[0056]

[0057] Calculate the first-order residual , This is the noise amplitude coefficient added for the first time.

[0058] Among them, targeting IMF extraction, Based on the first Adaptive noise is introduced into the first-order residual:

[0059] Extracted via EMD i IMF of order and ensemble average:

[0060] renew i Order residual:

[0061] The termination condition for adaptive decomposition is: when Decomposition stops when the function becomes monotonic or contains only low-frequency trend components.

[0062] In step 1022, CEEMDAN effectively improves the orthogonality and stability of each decomposition stage by introducing progressively adaptive amplitude noise. This process avoids mode aliasing in traditional EMD and eliminates residual noise in EEMD. In applications such as power load and photovoltaic power prediction, CEEMDAN can decompose the original signal into component sequences of multiple different frequency bands. This multi-scale decomposition mechanism helps subsequent prediction models capture and optimize features at each scale, thereby significantly improving overall prediction performance.

[0063] To reduce the influence of non-meteorological factors on the accuracy of load forecasting, the embodiments of the present application adopt the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method to decompose the original load sequence into multiple time scales. The method combines empirical mode decomposition (EMD) and ensemble average to obtain a set of intrinsic mode functions (IMF) and a residual term (Residue) by repeatedly superimposing Gaussian white noise with adaptive amplitude in the original sequence, thereby decomposing the original load sequence into components with different frequency characteristics. The ensemble average of the decomposition results is used to offset the randomness introduced by noise and obtain a stable set of IMF components.

[0064] Step 1023, determining an adaptive load type criterion based on the original meteorological data and the component sequences of the multiple different frequency bands.

[0065] Step 1023 includes: Calculating the correlation coefficients of the component sequences of the multiple different frequency bands and the meteorological variables in the original meteorological data; drawing a correlation coefficient variation curve based on the correlation coefficients; determining the inflection point where the correlation coefficient in the correlation coefficient variation curve decreases; and taking the inflection point as the adaptive load type criterion.

[0066] To obtain the meteorological load that best represents the meteorological sensitive characteristics, a comprehensive evaluation index S_ε is further constructed, which takes the clarity of the physical meaning of the meteorological load as the core. The index combines the "average correlation between the meteorological load and the meteorological variables" and the "stability of the decomposition results". The greater the average correlation, the better the synchronization between the decomposed meteorological load and the meteorological driving force, and the more explicit the physical meaning. The greater the stability index value (closer to 1), the less the decomposition results are affected by random noise, the higher the reproducibility, and the better the physical consistency. Through grid search experiments in the range of ε∈[0.05, 0.4], the optimal parameter is determined as the value of ε that maximizes S_ε. Under this parameter setting, the extracted meteorological load component has the highest correlation with key meteorological factors such as temperature and irradiance, and the decomposition process has good repeatability.

[0067] Step 1024, determining the meteorological load and the basic load based on the adaptive load type criterion and the component sequences of the multiple different frequency bands.

[0068] Step 1024 includes: Reconstructing the component sequences of the m frequency bands before the adaptive load type criterion into the meteorological load, and reconstructing the component sequences of the n frequency bands after the adaptive load type criterion and the residual term into the basic load.

[0069] For example, the component sequences of the m frequency bands before the adaptive load type criterion are high-frequency IMF components, and the component sequences of the n frequency bands after the adaptive load type criterion and the residual term are low-frequency IMF components.

[0070] After CEEMDAN decomposition, the high-frequency IMF component mainly reflects the load fluctuation caused by temperature, humidity, wind speed, irradiance and other meteorological factors, which is defined as meteorological load; the low-frequency IMF component and the residual term mainly represent the basic load composed of rigid demand such as daily life and industrial consumption.

[0071] In order to improve the applicability of CEEMDAN decomposition in meteorological load extraction, the key parameters of the embodiment are optimized. The noise amplitude coefficient ε is determined by grid search method, and the overall correlation between the high-frequency IMF component and the temperature sequence is maximized as the optimization objective; the integrated average number N is set based on the trade-off between decomposition stability and calculation efficiency.

[0072] In terms of the definition of decomposition layer number, the embodiment adopts the correlation coefficient method, the variance contribution rate method and the time-frequency characteristic observation method. By calculating the correlation coefficient of each IMF component and the key meteorological variable, and drawing its change curve, the inflection point of the first significant decline of the correlation coefficient is taken as the demarcation, which is used as the adaptive load type criterion. The first m IMF is reconstructed as meteorological load, and the remaining IMF and residual term are reconstructed as basic load. The comprehensive criterion can adaptively and rationally realize the physical meaning decoupling of load, and avoid relying on experience division.

[0073] In an embodiment, after separating out the meteorological load sequence, the correlation between the meteorological load sequence and external meteorological factors needs to be analyzed, so as to select the features with high value for prediction and input them into the prediction model for prediction. Figure 3 is a flowchart of step 103 provided by an embodiment of the present application, referring to Figure 3 Next, the detailed process of predicting the meteorological load prediction result according to the meteorological load and the original meteorological data by using the pre-constructed TimeMixer model will be described.

[0074] Therefore, step 103 includes: Step 1031, for any one meteorological feature, calculate the Pearson correlation coefficient between the meteorological feature and the meteorological load.

[0075] Wherein, the original meteorological data includes multiple meteorological features.

[0076] For example, the Pearson Correlation Coefficient (PCC) is used to quantify the degree of linear correlation between meteorological load and meteorological variables such as temperature, relative humidity, wind speed, and sunshine duration. The Pearson correlation coefficients between the meteorological load series and each candidate meteorological characteristic (such as temperature, humidity, wind speed, irradiance, cloud cover, etc.) are calculated to assess the strength and direction of the linear correlation. The resulting correlation coefficient matrix is ​​visualized using a heatmap to intuitively show the correlation between variables. Finally, the absolute value of the correlation coefficient (usually with a set threshold, such as |PCC|>0.6) and business knowledge are combined to screen out variables strongly correlated with the meteorological load.

[0077] The Pearson correlation coefficient measures the strength and direction of the linear relationship between two variables. The range of values ​​for the Pearson correlation coefficient can be set to [...]. [1,1], the closer to 1, the stronger the positive linear correlation; the closer to 1, the stronger the positive linear correlation. A value of 1 indicates a stronger negative linear correlation, while values ​​closer to 0 indicate a weaker linear relationship or no linear correlation at all. PCC is widely used in feature selection, dependency analysis, and regression modeling.

[0078] Given two random variables and The Pearson correlation coefficient is defined as:

[0079] In the formula, and They represent and The sample mean. If , and There is a positive correlation between the variables; if , and There is a negative correlation between the variables; if , and There is no correlation between the variables.

[0080] Step 1032: Select strongly correlated meteorological features based on Pearson correlation coefficient.

[0081] For example, the Pearson correlation coefficient ranges from [ [1,1], the closer to 1 or A value of 1 indicates a stronger positive / negative linear correlation, while values ​​closer to 0 indicate a weaker linear relationship or no linear correlation at all. Setting a threshold, such as 0.5, will retain only meteorological features with a correlation greater than 0.5 with the load as strongly correlated meteorological features.

[0082] Step 1033, input the weather load and strong related weather features into the pre-constructed TimeMixer model for targeted prediction, and output the weather load prediction result.

[0083] For example, the TimeMixer model adopts a full MLP architecture, including a multi-scale sequence generation module, a PDM (Past-Decomposable-Mixing, PDM) module, and a FMM (Future-Multipredictor-Mixing, FMM) module. The overall function flow is as shown in Figure 4 As shown, the time mixer structure involved is as shown in Figure 5 .

[0084] The multi-scale sequence generation module is used to divide the input weather load and original weather data into multiple scales to generate sequences representing short-term, medium-term, and long-term dynamics. The PDM module is used to extract features from each scale sequence to obtain scale information after feature extraction. The FMM module is used to integrate the scale information after feature extraction and output the weather load prediction result.

[0085] The time series of weather load and strong related weather features show significant differences under different sampling scales. Fine scales contain microscopic information such as short-term fluctuations, while coarse scales reflect macroscopic information such as long-term trends. To fully utilize this characteristic, in the multi-scale sequence generation module, the time series of input weather load and strong related weather features are transformed into multiple scales: Given the input sequence (P represents length, C represents variable quantity), generate M +1 scale sequences through average down-sampling. The sequence of the m th scale is represented as:

[0086] In the formula, is the original sequence (the finest scale), which retains the most detailed information; is the coarsest scale sequence, which highlights macro trend characteristics. After embedding layer processing, the multi-scale sequence extracts deep features to provide input data for subsequent mixing operations.

[0087] For example, the PDM module is used to extract key information from multi-scale historical sequences. Its core design is to separate seasonal and trend components and perform targeted mixing to adapt to their different characteristics. The seasonal component corresponds to short-term fluctuations, and the trend component corresponds to long-term changes. The specific process of extracting key information from multi-scale historical sequences is as follows: First, perform component decomposition: for each scale sequence (thel -1 layer input), using the sequence decomposition module in Autoformer (a long-term sequence prediction model based on deep decomposition architecture and self-correlation mechanism) (decomposition based on moving average) into seasonal component and trend component:

[0088] wherein, and denote the seasonal feature and trend feature of the l layer at the m scale, respectively, d_model is the feature dimension.

[0089] Then seasonal mixing and trend mixing are performed. The change of seasonal feature has the property that fine-scale patterns aggregate to form coarse-scale patterns (e.g. weekly cycles are composed of daily changes). Therefore, a bottom-up mixing strategy is adopted: starting from the finest scale (s m = 0), the detailed seasonal information of fine scale is transferred to coarse scale to complement the modeling of seasonal patterns of coarse scale.

[0090] Seasonal mixing (S-Mix) is realized through residual connection: = + Bottom-Up-Mixing(

[0091] wherein, Bottom-Up-Mixing(·) is a two-layer linear network containing GELU activation function, which maps the features of scale m -1 to scale m dimension.

[0092] The macro patterns of trend component are clearer on coarse scale, and the noise of fine scale may interfere with trend modeling. Therefore, a top-down mixing strategy is adopted: starting from the coarsest scale (s m = M ), the macro trend guides the trend learning of fine scale.

[0093] Trend mixing (T-Mix) is realized through residual connection: = + Top-Down-Mixing(

[0094] wherein, Top-Down-Mixing(·) is a two-layer linear network containing GELU activation function, which maps the features of scale m +1 to scale m dimension.

[0095] Finally, the features are integrated. The results of the seasonal mixing and the trend mixing are integrated by a feedforward network to obtain the output of the first layer:

[0096] The feedforward network contains linear transformations and GELU activations for cross-channel information interaction.

[0097] Exemplarily, after extracting key information from the multi-scale historical sequences, the FMM module is used to integrate the prediction capabilities of multi-scale features and generate the final prediction results of the future sequence. Since different scale sequences dominate different time series patterns (for example, fine scale is good at capturing short-term fluctuations, and coarse scale is good at capturing long-term trends), the prediction capabilities of the FMM module are complementary. The FMM module For the multi-scale features processed by the L-layer PDM, an independent predictor is designed for each scale:

[0098] wherein, (·) is a single-layer linear network that can directly regress the output sequence of length from the historical features of length after feature extraction.

[0099] The final meteorological load prediction result is obtained by fusing the complementary information of multiple scales by summation :

[0100] wherein, is the final output sequence, .

[0101] In view of the challenge of complex time series change prediction, the TimeMixer model proposed in the embodiment is a new type of time series prediction architecture, and the multi-layer structure thereof is composed of alternatingly stacked “token mixing” and “channel mixing” units. The token mixing unit realizes information exchange between different time steps (tokens) to capture long-term and short-term time series dependencies; the channel mixing unit integrates information of different feature channels to model the correlation between features. Each unit contains layer normalization (LayerNorm), residual connection and a nonlinear activation function (such as GELU), which effectively enhances the training stability and nonlinear representation capability. By alternately stacking these units, TimeMixer can simultaneously capture multi-scale time series dependencies and cross-feature interactions, thereby realizing high-precision multi-step prediction of meteorological load sequences. The core idea is to decouple the micro and macro patterns of time series through a multi-scale mixing mechanism and fuse the complementary prediction capabilities of multi-scale information.

[0102] The meteorological load sequence usually exhibits two key fluctuation characteristics: one is the instantaneous spike caused by weather mutations (such as cloud cover, gusts); the other is the smooth trend formed by the sustained meteorological conditions (such as heat waves, cold waves). Therefore, a single loss function is difficult to perform well on both modes. To this end, the present application designs a composite loss function to guide the TimeMixer model to learn these two characteristics evenly, in order to optimize the TimeMixer model's ability to capture the fluctuation characteristics of meteorological load. The composite loss function is composed of MAE, MAPE and gradient loss weighted. Among them, MAE ensures the overall trend to be consistent, MAPE strengthens the prediction accuracy of the high load period sensitive to meteorological conditions, and the gradient loss directly punishes the difference between the predicted curve and the true curve in the change trend, which is specifically used to improve the model's fitting ability for instantaneous meteorological fluctuations. Moreover, in order to optimize the model hyperparameters and evaluate its generalization performance, a forward rolling time series cross-validation strategy is adopted. By defining fixed training and validation windows and rolling them on the time axis, multiple training-validation data pairs conforming to the time series are generated. Taking the average validation performance of all rolling windows as the indicator, the optimal hyperparameter combination of TimeMixer is searched using Bayesian optimization. This method effectively avoids data leakage and ensures that the optimized model maintains stable prediction performance in different historical periods.

[0103] In an embodiment, to verify the effectiveness and practical applicability of the power load prediction method of the present application, the embodiment uses the power load and meteorological observation data of a certain region from 2020 to 2023 for three consecutive years for detailed case study. The original power load and original meteorological data are sampled at an hourly frequency, including power load measurement values and multi-dimensional meteorological variables such as irradiance, wind speed, wind direction, temperature, and relative humidity, which completely record the characteristic changes of the power system under different meteorological conditions, providing sufficient information support for feature extraction and prediction modeling.

[0104] For the prediction task, a direct prediction strategy is adopted: using 48 hours of historical load and meteorological data to directly predict the subsequent 24-hour load change curve. Compared with the indirect prediction method, this strategy can effectively avoid the accumulation of prediction errors and is more suitable for application scenarios with high requirements for short-term prediction accuracy. The model is implemented and trained based on the PyTorch deep learning framework, ensuring efficient training and inference of large-scale time series data.

[0105] The embodiment first uses the CEEMDAN algorithm to decompose the original power load sequence to extract multi-time scale feature information. After decomposition, three components are obtained, namely the basic load, the meteorological load and the residual term, as shown in Figure 6 The horizontal axis is time, with 540 points set within 48 hours, and the vertical axis unit is megawatts.

[0106] InFigure 6 In the middle, the first subgraph shows that the original power load curve fluctuates frequently, has no obvious periodicity, and has significant randomness and instability, making direct prediction difficult. After decomposition, the second subgraph presents a smooth basic load with clear periodicity, reflecting the basic electricity demand under the normal operation state of the power system. The third subgraph represents a meteorological load that fluctuates dramatically and is significantly affected by temperature, humidity, wind speed, and irradiance, making it a key research object for short-term load forecasting and the focus of the present invention's prediction modeling. The fourth subgraph shows that the residual term has a relatively small amplitude and is mainly random disturbance and noise, which has a weak influence on the overall prediction, so it is not considered in subsequent analysis.

[0107] To further explore the influence of meteorological factors on power load, the present embodiment uses Pearson correlation to quantify the correlation between meteorological load and each meteorological variable, and uses a correlation heat map for visualization, as shown in FIG. 3. The analysis shows that meteorological load is positively correlated with temperature, solar irradiance, and cloud cover, indicating that these factors have an important influence on load fluctuations; while the correlation with humidity, wind speed, wind direction, and precipitation is weak, among which meteorological load is moderately negatively correlated with wind direction. Based on the above findings, to improve the prediction performance of the model and avoid the interference of redundant / weakly correlated features, the present embodiment selects temperature, solar irradiance, and cloud cover, which have strong correlation, as the main input features for prediction. Figure 7

[0108] To verify the effectiveness of the proposed model in meteorological load prediction, two consecutive days of data were randomly selected from the test set as comparison samples, and compared with Autoformer, Informer, Fedformer, and LSTM-A, which are commonly used prediction models. The prediction curves of each model for the next 24 hours under the same conditions are shown in FIG. 4 (the horizontal axis time is time, with 96 points set for 24 hours; the vertical axis load value is load value, unit: megawatt), it can be seen that the fitting degree of the proposed model and the actual load (Real) curve is significantly better than that of the comparison models, and its prediction result can more accurately track the load trend and capture the fluctuation details. Although other models can also generate prediction results, they have insufficient fidelity at the peak and trough, with obvious deviations. Among them, the LSTM-A model has the largest prediction error, showing obvious lag and cumulative bias. Figure 8 The root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are used as evaluation criteria, and the quantitative evaluation indicators of each model on the test set are shown in Table 1.

[0109] Table 1. Prediction accuracy of each model

[0110] ​​

[0111] Autoformer, Informer and Fedformer perform similarly in short-term prediction, with RMSE of 1.6804, 1.7254 and 1.7711, and MAPE of 6.84%, 7.29% and 7.31% respectively. Although these models can better capture the overall trend, they still have deficiencies in the details of fluctuations.

[0112] The RMSE and MAPE of the LSTM-A model are 1.9178 and 7.84%, respectively, which is inferior to the Transformer-based model. The prediction curve is less smooth, and there is a large error in the peak and fluctuation section.

[0113] The proposed model achieves the optimal results of RMSE 1.5216, MAE 1.2386 and MAPE 6.32%, which are better than the comparative models, showing its excellent ability in improving the accuracy of short-term meteorological load prediction In summary, the superiority of the proposed model is confirmed through the comparison of visual curves and quantitative index analysis, effectively verifying its effectiveness and applicability in meteorological load prediction tasks.

[0114] It can be seen that the power load prediction method proposed by the present application has higher accuracy and fitting degree compared with the traditional algorithm based method, can realize high-quality power load prediction, improve the accuracy of load prediction, help the healthy development of the power market, and promote the overall progress of the power industry.

[0115] The CEEMDAN algorithm is used to adaptively decompose the preprocessed power load sequence to obtain a plurality of intrinsic mode functions (IMFs) of different frequencies. The components are reconstructed according to the physical meaning: the high-frequency IMF is defined as the'meteorological load', representing the random fluctuations driven by meteorological factors; the low-frequency IMF and the residual term are reconstructed as the 'base load', representing the stable base load formed by the laws of social production and life. This step realizes the physical decoupling of the complex load signal and lays the foundation for subsequent targeted modeling. Secondly, for the'meteorological load' component, meteorological correlation analysis and feature selection are carried out. By calculating the Pearson correlation coefficients of temperature, humidity, irradiance, wind speed and other meteorological variables, and supplemented by heat map visualization, the key meteorological features are selected as model inputs to reduce redundancy and improve prediction efficiency. Subsequently, the TimeMixer model is constructed and trained for targeted prediction. The selected meteorological features and historical meteorological load data are input into the model, the multi-scale sequence generation module is used to capture short-term fluctuations and long-term trends, the PDM module is used to mix seasonal and trend components respectively, and the FMM module is used to integrate multi-scale information to output accurate prediction of the meteorological load. In the entire prediction process, CEEMDAN decomposition provides a target pure input for TimeMixer, and TimeMixer realizes accurate fitting of the meteorological load by virtue of its multi-scale mixing mechanism. The two are deeply coordinated to form an end-to-end optimized prediction link. Finally, the predicted meteorological load and the base load are superimposed and reconstructed to obtain a high-precision day-ahead total power load prediction value, thereby providing reliable support for power system dispatching decisions. The process logic is clear, the technologies in each link are closely coordinated, and both physical interpretability and prediction accuracy are considered.

[0116] Referring to Figure 9 One embodiment provides a controller for performing the power load prediction method in the above embodiment.

[0117] The embodiments of the present application also provide a controller, referring to Figure 9 The controller 300 can include at least one processor 310 and a memory 320, the memory 320 storing a computer program executable on the at least one processor 310, and the processor 310 implements the steps in any of the above method embodiments when executing the computer program.

[0118] For example, the computer program can be divided into one or more modules / units, which are stored in the memory 320 and executed by the processor 310 to complete the present application. One or more modules / units can be a series of computer program segments that can complete a specific function, which are used to describe the execution process of the computer program in the controller 300.

[0119] Those skilled in the art can understand that,Figure 9 The controller is merely an example and does not constitute a limitation on the controller, and can include more or fewer components than shown, or combine certain components, or have different components, such as input / output devices, network access devices, buses, and the like.

[0120] The processor 310 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, or the like. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0121] The memory 320 can be an internal storage unit of the controller, or an external storage device of the controller, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, or the like. The memory 320 is used to store computer programs and other programs and data required by the controller. The memory 320 can also be used to temporarily store data that has been output or will be output.

[0122] The bus can be an industry standard architecture (ISA) bus, a peripheral component (PCI) bus, an extended industry standard architecture (EISA) bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, the bus in the drawings of the present application does not limit to only one bus or one type of bus.

[0123] The power load prediction method provided by the embodiments of the present application can be applied to a controller such as a computer, a tablet computer, a notebook computer, a netbook, a personal digital assistant (PDA), and the like. The embodiments of the present application do not make any limitation on the specific type of the controller.

[0124] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0125] In addition, although the operations of the method of the present application are described in a specific order in the accompanying drawings, this does not require or imply that the operations must be performed in this specific order, or that all of the shown operations must be performed to achieve the desired result. Additionally or alternatively, certain steps can be omitted, a plurality of steps can be combined into one step, and / or one step can be divided into a plurality of steps.

[0126] The present application also provides a computer readable storage medium, the computer readable storage medium stores computer execution instructions, when the processor executes the computer execution instructions, the power load prediction method provided by the above embodiments of the present application is realized.

[0127] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in a certain embodiment can be referred to the relevant description of other embodiments.

[0128] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for predicting electricity load, characterized in that, include Acquire raw power load and raw meteorological data; each raw power load corresponds one-to-one with each raw meteorological data point. The original power load is adaptively decomposed based on the original meteorological data to obtain meteorological load and base load; the meteorological load represents the random fluctuation load driven by meteorological factors. Using a pre-built TimeMixer model, the meteorological load prediction result is obtained based on the meteorological load and the original meteorological data; The meteorological load forecast results are superimposed and reconstructed with the base load to obtain the day-ahead power load forecast results.

2. The power load forecasting method as described in claim 1, characterized in that, The adaptive decomposition of the original power load based on the original meteorological data to obtain meteorological load and base load includes: The original power load is preprocessed to obtain the preprocessed power load; The preprocessed power load is adaptively decomposed to obtain component sequences of multiple different frequency bands; The adaptive load type criterion is determined based on the original meteorological data and the component sequences of the multiple different frequency bands; The meteorological load and the base load are determined based on the adaptive load type criterion and the component sequences of the multiple different frequency bands.

3. The power load forecasting method as described in claim 2, characterized in that, The determination of adaptive load type criteria based on the original meteorological data and the component sequences of multiple different frequency bands includes: Calculate the correlation coefficients between the component sequences of each different frequency band and the meteorological variables in the original meteorological data; Plot the correlation coefficient variation curve based on the aforementioned correlation coefficient; Identify the inflection point where the correlation coefficient decreases in the correlation coefficient change curve; use the inflection point as the criterion for the adaptive load type.

4. The power load forecasting method as described in claim 3, characterized in that, The determination of the meteorological load and the base load based on the adaptive load type criterion and the component sequences of the multiple different frequency bands includes: The component sequences of the first m frequency bands of the adaptive load type criterion are reconstructed into meteorological load, and the component sequences of the last n frequency bands of the adaptive load type criterion and the residual terms are reconstructed into basic load.

5. The power load forecasting method as described in claim 1, characterized in that, The process of using a pre-built TimeMixer model to predict meteorological load based on the meteorological load and the original meteorological data to obtain the meteorological load prediction result includes: For any given meteorological feature, calculate the Pearson correlation coefficient between that meteorological feature and the meteorological load; wherein the original meteorological data includes multiple meteorological features; Strongly correlated meteorological features were selected based on the Pearson correlation coefficient. The meteorological load and the strongly correlated meteorological features are input into a pre-built TimeMixer model for targeted prediction, and the meteorological load prediction result is output.

6. The power load forecasting method as described in claim 2, characterized in that, The adaptive decomposition of the preprocessed power load to obtain component sequences of multiple different frequency bands includes: The preprocessed power load is adaptively decomposed using the CEEMDAN algorithm to obtain component sequences of multiple different frequency bands.

7. The power load forecasting method as described in claim 6, characterized in that, The preprocessed power load is adaptively decomposed using the CEEMDAN algorithm to obtain component sequences in multiple different frequency bands, including: The original power load sequence is determined based on the preprocessed power load. Gaussian white noise of different amplitudes is superimposed on the original power load sequence to construct multiple noise-assisted sequences; Each set of noise-assisted sequences is decomposed using CEEMD to obtain the decomposition results; each decomposition result includes the corresponding IMF component. The corresponding IMF components are integrated and averaged to obtain the IMF component set.

8. The power load forecasting method as described in claim 1, characterized in that, The TimeMixer model includes a multi-scale sequence generation module, a past decomposition and mixing module, and a future multi-prediction and mixing module. The multi-scale sequence generation module is used to divide the input meteorological load and the original meteorological data into multiple scales to generate sequences at each scale that represent short-term, medium-term and long-term dynamics respectively. The past decomposition and mixing module is used to extract features from sequences at each scale to obtain information at each scale after feature extraction. The future multi-prediction hybrid module is used to integrate information from various scales after feature extraction and output meteorological load prediction results.

9. The power load forecasting method as described in claim 4, characterized in that, The component sequences of the first m frequency bands of the adaptive load type criterion are high-frequency IMF components; the component sequences of the last n frequency bands of the adaptive load type criterion and the residual term are low-frequency IMF components.

10. The power load forecasting method as described in claim 2, characterized in that, The preprocessing of the original power load includes: The original power load was processed using Z-score standardization; The missing data of the original power load were filled in using linear interpolation.

Images