Multi-time scale power load accurate prediction method, device, equipment and medium

Abstract

The application provides a multi-time-scale power load accurate prediction method, device, equipment and medium, relates to the power system prediction technical field, and the method is first to pretreat historical power load data, adopts a complete set of empirical mode decomposition algorithm to adaptively decompose the original load sequence into intrinsic mode components and residual components, and is divided into high, medium and low frequency component groups according to the spectral characteristics.Further introduce the cross-frequency band gating fusion mechanism, through the gating control unit adaptive learning the mutual influence relationship between different frequency band information, make the low frequency trend constraint high frequency fluctuation, the medium frequency periodicity guide high frequency peak and valley change, use high frequency disturbance to correct the medium and low frequency characteristics, generate a single comprehensive load sequence.The sequence is input into a multi-time-scale prediction model, and a quantile regression method is introduced to output a prediction confidence interval.The application can significantly improve the accuracy, stability and credibility of the power load prediction result.

Description

Technical Field

[0001] This invention relates to the field of power system forecasting technology, specifically to methods, devices, equipment, and media for accurate forecasting of power load at multiple time scales. Background Technology

[0002] Electricity load forecasting is a crucial technical aspect of modern power system operation and planning. Its accuracy directly impacts the economic dispatch, safe and stable operation of the power grid, and the trading efficiency of the electricity market. With the continuous advancement of new power system construction, load forecasting's strategic importance in ensuring power grid supply and demand balance, improving energy efficiency, and promoting the digital transformation of the energy system is increasingly prominent. Electricity load sequences are inherently typical non-stationary time series, their dynamic changes influenced by a complex coupling of multiple factors. These include meteorological factors such as temperature, humidity, and wind speed significantly affecting temperature-sensitive loads; daily, weekly, and seasonal periodicity; and nonlinearity and uncertainty brought about by social factors such as macroeconomic fluctuations, industrial structure evolution, holiday arrangements, and unforeseen events. This multi-scale, strongly nonlinear characteristic makes it difficult to accurately characterize load variation patterns using simple models.

[0003] Currently, commonly used load forecasting methods in power systems still have certain limitations. Traditional time series models based on statistical theory, such as ARIMA (Autoregressive Integrated Moving Average) and its improved models, while possessing a relatively sound mathematical foundation, have limited ability to characterize complex nonlinear relationships. Regression models based on machine learning, such as Support Vector Machines and Random Forests, have improved prediction accuracy to some extent, but their ability to model long-term time series dependencies is insufficient, and they heavily rely on manual feature engineering and empirical settings. In recent years, deep learning methods have been gradually introduced into the field of power load forecasting. However, single deep learning models such as LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), or Transformer often struggle to simultaneously consider both local detailed features and global time dependencies in the load sequence. For example, LSTM exhibits a lag in its response to sudden load fluctuations, while Transformer is relatively insensitive to subtle local changes. More importantly, existing methods typically use non-stationary raw load sequences directly as model input, failing to fully exploit their inherent multi-timescale structural characteristics. This limits the model's adaptability and generalization performance to complex load change patterns. Related industry reports indicate that when facing highly volatile load curves, existing forecasting methods generally exhibit high Mean Absolute Percentage Error (MAPE), which is further amplified in special scenarios such as extreme weather. This not only affects grid operation safety but also significantly increases system operating costs.

[0004] To address the aforementioned issues, existing research primarily focuses on improvements in two areas: signal decomposition and model fusion. On one hand, signal decomposition methods such as wavelet transform and empirical mode decomposition are used to perform multi-scale analysis of load sequences. However, traditional decomposition methods are prone to mode aliasing and endpoint effects, are highly sensitive to parameter settings and noise, and lack engineering stability. On the other hand, model integration or fusion strategies are used to improve prediction performance. However, simply stacking models often leads to high computational complexity and lacks effective feature interaction mechanisms, which can easily cause overfitting and limit its engineering applicability.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention provides a method, apparatus, equipment, and medium for accurate prediction of power load at multiple time scales, which can at least partially improve the above-mentioned problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for accurate prediction of power load across multiple time scales, comprising: Historical power load data is acquired, and the historical power load data is preprocessed and decomposed to obtain residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components, and low-frequency intrinsic mode components. The residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components, and low-frequency intrinsic mode components are input into a preset cross-band gated fusion module. The mutual constraint relationship between the intrinsic mode components of different frequency bands is adaptively learned to achieve collaborative modeling and dynamic fusion of multi-band information, and the intrinsic mode components of different frequency bands are fused into a single comprehensive load sequence. The single comprehensive load sequence is input into a pre-trained multi-timescale prediction model for prediction processing to generate the final load prediction sequence. The quantile regression method is used to estimate the confidence interval of the final load forecast sequence, and the predicted values ​​and prediction confidence intervals of multiple quantiles are calculated. Post-processing and prediction performance evaluation are performed on the predicted values ​​and prediction confidence intervals of multiple quantiles to obtain a visualization report.

[0009] The present invention also provides a multi-timescale power load accurate prediction device, which includes: The decomposition unit is used to acquire historical power load data, perform data preprocessing and signal decomposition on the historical power load data, and obtain residual components, high-frequency intrinsic mode component groups, mid-frequency intrinsic mode component groups and low-frequency intrinsic mode component groups. The fusion unit is used to input the residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components and low-frequency intrinsic mode components into a preset cross-band gated fusion module. The module adaptively learns the mutual constraint relationship between the intrinsic mode components of different frequency bands to achieve collaborative modeling and dynamic fusion of multi-frequency band information, and fuses the intrinsic mode components of different frequency bands into a single comprehensive load sequence. The prediction unit is used to input a single comprehensive load sequence into a pre-trained multi-timescale prediction model for prediction processing, and generate the final load prediction sequence. The estimation unit is used to perform confidence interval estimation on the final load forecast sequence using the quantile regression method, and calculates the predicted values ​​and prediction confidence intervals for multiple quantiles. The evaluation unit is used to perform post-processing and prediction performance evaluation on the predicted values ​​and prediction confidence intervals of multiple quantiles, and to obtain a visual report.

[0010] The present invention also provides a device for accurate prediction of power load at multiple time scales, which includes a memory and a processor. The memory stores a computer program that can be executed by the processor to implement the method for accurate prediction of power load at multiple time scales as described above.

[0011] The present invention also provides a computer-readable storage medium, characterized in that it stores a computer program, which can be executed by a processor of the device in which the computer-readable storage medium is located, to realize the multi-time-scale accurate power load prediction method as described in any of the above claims.

[0012] In summary, given the insufficient prediction accuracy and reliability of existing power load forecasting methods when dealing with non-stationary, multi-scale load sequences, this invention proposes a high-precision power load forecasting scheme based on signal decomposition and hybrid deep learning. This method, by introducing optimization algorithms and uncertainty modeling mechanisms, achieves precise characterization and reliable prediction of complex load variation characteristics.

[0013] Compared with existing technologies, this invention has the following advantages: It effectively handles the non-stationary characteristics of power load sequences through CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) decomposition, decoupling the original load signal into high, medium, and low frequency IMF (Intrinsic Mode Function) components and residual components (RES). Furthermore, it introduces a cross-frequency gated fusion (CF-GF) mechanism to achieve collaborative construction of information from different frequency bands. This invention integrates model and dynamics; it adaptively tunes the hyperparameters of the prediction models corresponding to each component using optimization algorithms, extracts local features using a TCN (Temporal Convolutional Network), performs deep temporal modeling using an ABiLSTM (Attention-based Bidirectional Long Short-Term Memory) architecture, and finally fuses the multi-branch prediction results using a Seq2Seq (Sequence to Sequence) model to generate accurate load prediction values. Simultaneously, it constructs prediction confidence intervals based on quantile regression. This overall approach effectively improves the accuracy, stability, and interpretability of load prediction. Experiments have demonstrated that this method exhibits excellent performance on load data from different seasons and regions, with the average absolute percentage error consistently controlled within 3%, providing reliable technical support for power system dispatching and market transactions. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the multi-timescale power load accurate prediction method provided in the first embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the multi-timescale power load accurate prediction device provided in the second embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] refer to Figure 1As shown, the first embodiment of the present invention discloses a method for accurate prediction of power load at multiple time scales, which can be executed by a device for accurate prediction of power load at multiple time scales (hereinafter referred to as the prediction device), specifically, by one or more processors within the prediction device, to implement the following method: S1. Acquire historical power load data, perform data preprocessing and signal decomposition on the historical power load data to obtain residual components, high-frequency intrinsic mode component groups, mid-frequency intrinsic mode component groups and low-frequency intrinsic mode component groups. Specifically, step S1 further includes: acquiring historical power load data and performing data preprocessing on the historical power load data, wherein the data preprocessing includes: missing value repair preprocessing, outlier removal preprocessing, and normalization preprocessing. The preprocessed historical power load data is adaptively decomposed using a complete ensemble empirical mode decomposition algorithm to obtain intrinsic mode component sets (IMF1, IMF2, ..., IMF). n The intrinsic modal components are (and the residual components (long-term variation trend)) and the intrinsic modal components are arranged in descending order of frequency. The intrinsic mode components are input into the preset IMF-GAN adversarial noise reduction module for noise reduction, resulting in a clean multi-band load subsequence. Based on the spectral characteristics and energy distribution of the multi-band load subsequence, the multi-band load subsequence is divided into high-frequency intrinsic mode component group, mid-frequency intrinsic mode component group and low-frequency intrinsic mode component group.

[0018] In this embodiment, historical power load data is first acquired. This data typically originates from the power grid dispatch system or smart meter data acquisition devices, with a time granularity of 15 minutes, 1 hour, or 1 day, determined based on the actual forecasting task. After acquiring the raw load data, comprehensive data preprocessing is performed. Preprocessing includes three operations: First, missing value repair preprocessing: for missing data points caused by communication failures or acquisition anomalies, linear interpolation, cubic spline interpolation, or the average load at the same time on adjacent days can be used to fill in the missing data to ensure the continuity of the time series. Second, outlier removal preprocessing: for peaks or troughs that significantly deviate from the normal load range, such as values ​​much higher or lower than historical levels due to equipment false alarms, methods based on box plots or the three sigma rule are used to identify and remove them. After removal, interpolation is used to supplement the missing data. Third, normalization preprocessing: to eliminate the influence of different units and numerical ranges on the training of the deep learning model, maximum-minimum normalization or Z-score normalization methods are used to map all load data to a unified numerical range, such as between 0 and 1. Through the above preprocessing, the quality of the original load data is significantly improved, laying a clean, continuous, and standardized data foundation for subsequent decomposition and modeling.

[0019] After data preprocessing, a complete ensemble empirical mode decomposition algorithm is used to adaptively decompose the preprocessed historical power load data. Unlike traditional empirical mode decomposition or ensemble empirical mode decomposition, complete ensemble empirical mode decomposition introduces paired positive and negative Gaussian white noise with adaptive amplitude and calculates the local mean of the residual signal in multiple iterations. This enables it to adaptively decompose the non-stationary, nonlinear original load sequence into a set of intrinsic mode components and a residual component with fewer iterations and lower reconstruction error. All the decomposed intrinsic mode components are arranged in descending order of frequency. High-frequency components mainly correspond to random load fluctuations and noise components, mid-frequency components reflect regular electricity consumption patterns such as daily and weekly cycles, low-frequency components reflect long-term trend changes, and the residual component represents the overall trend of the load sequence, such as slow changes under seasonality or macroeconomic influences. Through complete set empirical mode decomposition, the original complex signal is effectively decoupled, and each component is clearer in a physical sense. This provides a structured data foundation for subsequent differentiated processing of different frequency band characteristics, while effectively suppressing mode aliasing and avoiding the accumulation of reconstruction errors.

[0020] In simple terms, the CEEMDAN decomposition process involves repeatedly introducing pairs of positive and negative Gaussian white noise with zero mean and adaptive amplitude. This noise is then superimposed on the original load signal before Empirical Mode Decomposition (EMD) is performed to reduce mode aliasing and improve decomposition stability. During the extraction of each Intrinsic Mode Component (IMF), the difference between the current residual signal and its local mean envelope is calculated iteratively until the residual signal exhibits a monotonic trend. This adaptively and completely decomposes the original load signal into a set of IMF components and residual components (RES) arranged from high to low frequency.

[0021] To further improve the purity of each intrinsic mode component, this invention inputs all the intrinsic mode components obtained from decomposition into a preset IMF-GAN adversarial denoising module for denoising processing. Specifically, the generator fits the real load fluctuation characteristics, while the discriminator identifies and removes high-frequency noise, pseudo-modes, and spurious fluctuations. This achieves component-level precise denoising while fully preserving key details such as load abrupt changes and spikes, outputting a clean multi-band load subsequence and improving the stability of subsequent fusion and prediction. This module employs a generative adversarial network architecture. The generator takes the noisy intrinsic mode components as input and gradually expands the receptive field through multi-layered stacked dilated causal convolutions, learning and fitting the distribution characteristics of the real effective load fluctuations in the component, and outputting denoised candidate components. The discriminator simultaneously receives the original noise-free real components and the denoised components output by the generator, using a binary classification cross-entropy loss function to determine whether the input is a real effective feature or residual noise. During adversarial training, the generator and discriminator engage in a game of mutual competition and alternating optimization, ultimately enabling the generator to effectively remove high-frequency random noise, pseudo-modes, and spurious fluctuations without destroying key engineering features such as load abrupt changes, peak values, and sudden drops. After processing by the IMF-GAN adversarial denoising module, a clean multi-band load subsequence is obtained. Compared to traditional filtering or wavelet soft thresholding methods, this denoising method achieves high-fidelity denoising at the component level, significantly improving the signal-to-noise ratio while preserving key load details to the maximum extent. This avoids feature loss or waveform distortion caused by excessive smoothing, providing high-quality and reliable input data for subsequent frequency band segmentation and fusion prediction.

[0022] Specifically, the IMF-GAN adversarial denoising and reconstruction consists of an adversarial training architecture comprised of a generator network and a discriminator network. The generator employs a multi-layer stacked dilated causal convolution, with the convolution kernels using an exponentially increasing dilation factor to progressively expand the temporal receptive field. This allows it to fully capture fine temporal features within the IMF components, such as short-term abrupt changes, local spikes, continuous fluctuations, and gradual trend changes. It fits the true data distribution of the load components through forward propagation, outputting denoised candidate components that retain all effective details. The discriminator uses a binary classification network structure combining convolution and fully connected layers. It uses the original, noise-free, true IMF components as positive samples and components containing high-frequency noise / pseudo-modalities / false fluctuations as negative samples. For negative samples, the discrimination probability is calculated using binary cross-entropy loss to accurately distinguish between real effective load features and redundant noise components. During adversarial training, the generator continuously optimizes the denoising effect with the goal of deceiving the discriminator, while the discriminator strengthens its noise discrimination ability with the goal of improving the recognition accuracy. The two converge dynamically through a minimax game, maximizing the removal of random noise, spurious modes, and abnormal disturbances while strictly preserving key engineering features such as load mutation points, peak values, sudden drops, and inflection points. This fundamentally avoids the defects of traditional filtering algorithms, such as feature loss, waveform distortion, phase shift, and detail smoothing, achieving distortion-free, high-precision, and high-fidelity adversarial denoising at the IMF component level.

[0023] After obtaining the purified multi-band load subsequence, the pure intrinsic mode components are automatically divided into high-frequency intrinsic mode component groups, mid-frequency intrinsic mode component groups, and low-frequency intrinsic mode component groups based on the spectral characteristics and energy distribution of each subsequence. This constructs a multi-band load subsequence, thereby decoupling the original non-stationary and nonlinear load signal into multi-scale components with clearer structural characteristics. Specific division criteria can include the center frequency, energy proportion, or correlation coefficient with the original load sequence. For example, components with a center frequency higher than a preset threshold and an energy proportion lower than a certain threshold are classified as high-frequency components, mainly reflecting instantaneous load fluctuations and noise components; components with obvious diurnal or weekly periodic characteristics and relatively concentrated energy are classified as mid-frequency components, corresponding to regular electricity consumption behavior; and components with lower frequencies, gradual changes, and a larger energy proportion are classified as low-frequency components, reflecting the basic level and long-term evolution trend of the load. Through this division, the original non-stationary and nonlinear load signal is successfully decoupled into multi-scale component groups with clearer structural characteristics, namely, high-frequency, mid-frequency, and low-frequency groups, as well as a separately retained residual component. This division method enables subsequent modeling to adopt differentiated processing strategies based on the physical characteristics of different frequency bands, avoiding the mutual interference problem caused by mixing multi-scale information in the same model, thereby improving the pertinence and effectiveness of the overall prediction framework.

[0024] It is important to note that during this process, residual components are retained as long-term trend signals. This decomposition method effectively suppresses the mode mixing problem in traditional EMD methods, while avoiding the reconstruction error introduced by EEMD methods, providing a reliable foundation for subsequent cross-band information fusion and multi-timescale prediction.

[0025] S2, the residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components, and low-frequency intrinsic mode components are input into a preset cross-band gated fusion module. Here, the mutual constraints between the intrinsic mode components of different frequency bands are adaptively learned to achieve collaborative modeling and dynamic fusion of multi-band information, fusing the intrinsic mode components of different frequency bands into a single comprehensive load sequence. This fused sequence retains key information from each frequency band while suppressing noise and redundant interference, providing a stable and accurate input for subsequent predictions. Specifically, this comprehensive load sequence maintains consistency with the original load sequence in the time dimension, but incorporates high-frequency fluctuation characteristics, mid-frequency periodic characteristics, and low-frequency trend information in its numerical characteristics. This suppresses noise and redundant information while retaining key characteristics of multi-band load changes, serving as a unified input for subsequent multi-timescale predictions.

[0026] Specifically, step S2 further includes: using the load operation state determination unit to determine the high-frequency intrinsic mode component group, the mid-frequency intrinsic mode component group, and the low-frequency intrinsic mode component group, classifying the current load state into a steady state, a volatile dynamic state, an abrupt change state, and a periodically dominant state. A three-component Gaussian mixture model is used to fit the probability distribution of the fluctuation intensity sequence to construct a mixture distribution. Three Gaussian distributions were obtained. , For a steady state, It is a periodic dominant state. The state is abruptly changed, where V is the wave intensity and k is the Gaussian component index. The mixing weight coefficients for the k-th Gaussian component are... It follows a Gaussian distribution. Let be the mean of the k-th Gaussian component. Let be the variance of the k-th Gaussian component; Based on the current load status, corresponding paths are assigned, and feature enhancement is performed on the corresponding frequency band within each path. Cross-path collaborative fusion is carried out at the exit of all paths to form a preliminary fusion result. Specifically, when the load status is in a stable state, low-frequency components enter the trend path, mid-frequency components weakly participate in the periodic path, and high-frequency components are suppressed. When the load status is in a periodic dominant state, mid-frequency components enter the periodic path, low-frequency components enter the trend path, and high-frequency components enter the abrupt path. When the load status is in an abrupt state, high-frequency components enter the abrupt path, mid-frequency components enter the periodic path, and low-frequency components enter the trend path.

[0027] In this embodiment, a load operation status determination unit is introduced. Based on multi-dimensional characteristics (fluctuation intensity, gradient change rate, periodic consistency, anomaly index, etc.), a state discrimination function is constructed to classify the current load status into steady state, fluctuating state, abrupt change state, periodically dominant state, etc., and different cross-frequency band fusion strategies are triggered. First, the normalized fluctuation intensity at time t is set. for: , The original load at time t, The original load at time t-1, Let be the overall fluctuation scale (standard deviation). The standard deviation is taken as the sliding window standard deviation, and the formula for calculating the standard deviation at time t is: The meaning of this formula is: using the most recent The fluctuation scale is calculated at each point and adaptively varied. Then, the gradient rate of change at time t is set. for: Furthermore, the gradient rate of change can be used to determine whether a state change is slow or abrupt. for The first derivative.

[0028] The load status can be determined based on the set normalized fluctuation intensity and gradient rate of change, as shown in Table 1. Table 1:

[0029] Among them, the first threshold Second threshold The value of can be obtained through a Gaussian mixture model (GMM). Specifically, this method uses a three-component Gaussian mixture model to fit the probability distribution of the fluctuation intensity sequence and construct a mixture distribution; therefore, the first threshold... Second threshold The value of is: =intersection point , =intersection point Simply put, the first threshold Second threshold The first threshold was obtained by statistically analyzing the distribution of historical load change rates. The second threshold is the threshold for the lower and middle quantiles of the rate of change. Using a high quantile threshold for the rate of change, adaptive segmentation for different load states can be achieved. The corresponding fusion strategies are shown in Table 2; Table 2:

[0030] In summary, the V and G hard thresholds are merely empirical classifications, using experience to determine the size of the hard threshold. The generation of the three Gaussian distributions, however, is adaptive, automatically learning two boundary thresholds from the data, and then using these two boundary thresholds to further divide the three states. Therefore, V and G are judgment indicators, while GMM is used to automatically generate the judgment thresholds, providing automated boundaries.

[0031] Secondly, a cross-frequency band information routing network is constructed. Based on the current load status, high-frequency, mid-frequency, and low-frequency components are dynamically allocated to different feature channels. The information propagation path is determined by a path selection function, achieving cross-frequency band asymmetric information interaction. This upgrades "weighted fusion" to "structure-level routing fusion." Its core logic is that features from different frequency bands (high-frequency = abrupt changes, mid-frequency = periodic information, low-frequency = trend information) are no longer mixed into the same channel. Instead, they are dynamically allocated to different feature propagation paths based on state-triggered commands, achieving information decoupling before targeted fusion—a more advanced and targeted approach than traditional Attention weighting. This step constructs three core feature paths: a short-term abrupt change path (adapted to high-frequency), a periodic regularity path (adapted to mid-frequency), and a long-term trend path (adapted to low-frequency). In actual use, the path is allocated based on state-triggered commands. Specifically, when the load is in a stable state, "low-frequency dominant routing" is triggered—low-frequency components enter the trend path (main path), mid-frequency components weakly participate in the periodic path (providing phase constraints), and high-frequency components are suppressed (not entering any path). When the load is in a periodic dominant state, "mid-frequency dominant routing" is triggered—mid-frequency components enter the periodic path (main path), low-frequency components enter the trend path (providing trend anchoring), and high-frequency components enter the abrupt change path (providing phase constraints). When the load is in an abrupt change state, "high-frequency priority routing" is triggered—high-frequency components enter the abrupt change path (main path), mid-frequency components enter the periodic path (providing phase constraints), and low-frequency components enter the trend path (providing trend anchoring). It should be noted that the mixed state is handled by: three-path parallel activation + adaptive gating automatic weighting. The mixed state can be directly fused by model adaptation without designing a new path separately.

[0032] Subsequently, feature enhancement is performed on the corresponding frequency band within each path (e.g., enhancing the high-frequency mutation capture capability of a mutation path). Cross-path collaborative fusion is performed at the path exit to form a preliminary fusion result. The "fusion direction" triggered by the state provides "preliminary fusion results to be optimized" for consistency constraints. At the same time, the logic of routing path allocation will also adapt to the requirements of the constraints (e.g., the output of the trend path must meet the trend consistency constraint).

[0033] The core objective of the trend path is to enhance the trend stability of low-frequency components, filter out irrelevant noise, provide a long-term trend benchmark for the fusion results, and adapt to all load conditions. It employs a trend smoothing enhancement formula to suppress minor noise in the low-frequency components while preserving the core trend, resulting in the core formula for the trend path: , A trend preservation coefficient, with a value range of [0.7, 0.9], effectively preserves low-frequency trends. This is a moving average filter for the low-frequency components, where u is the moving window length, which is consistent with the window length for the standard deviation mentioned earlier. Used to smooth noise; in steady state In the altered state , It is a low-frequency component.

[0034] The core objective of the periodic path is to enhance the periodicity of the mid-frequency components, extract peak-valley phase information, provide periodic guidance for the fusion results, and avoid phase shifts caused by high-frequency abrupt changes. Its core formula is: , For mid-frequency components, This is the periodic enhancement coefficient, with a value range of [0.5, 0.8], which enhances the peak-valley difference. The peak value of the intermediate frequency component. This represents the valley value of the mid-frequency component. The mean of the mid-frequency components; during the periodic dominant state. In steady state .

[0035] The core objective of the mutation path is to enhance the mutation characteristics of high-frequency components, suppress high-frequency noise, accurately capture short-term sudden changes in load, and avoid excessive amplification of mutation information. Its core formula is: , For high-frequency components, This is the maximum absolute value of the high-frequency components, used for normalization to avoid numerical overflow. This is the mutation gain coefficient, with a value range of [0.8, 1.2], which controls the mutation intensity. The sign function is used for high-frequency components, preserving the mutation direction (positive mutation / negative mutation); in the mutation state... In steady state .

[0036] Finally, after obtaining the preliminary fusion results, RES is directly superimposed for trend-based compensation, and cross-frequency consistency constraint fusion processing is performed to generate the final single integrated load sequence. Specifically, based on dynamic routing fusion, mathematical constraints are added to force the fusion results to meet the three consistency criteria of "trend, phase, and energy," preventing the fusion results from deviating from the load time series pattern, while further optimizing the fusion weights.

[0037] The trend consistency constraint refers to the low-frequency trend component of the fusion result, which must maintain consistency with the trend of the original low-frequency component. The formula is as follows: , The original low-frequency components are used to avoid trend shifts after route fusion (e.g., in a steady state, the fusion result needs to conform to the low-frequency trend). Phase consistency constraints require that the peak and valley positions of the fusion result be aligned with the periodic peak and valley positions of the mid-frequency components to avoid peak and valley shifts caused by high-frequency abrupt changes (e.g., in abrupt changes, high frequencies dominate but must follow the periodic phase of the mid-frequency components). Energy conservation constraints define the total energy of the fusion result as the sum of the energies of each frequency band component, expressed by the formula: , Let be the energy of the i-th frequency band component (low frequency, mid frequency, high frequency); avoid information loss or over-amplification (for example, when allocating routes, the energy proportion of each path must meet this constraint). Constraint optimization fusion is an optimization objective that incorporates the three constraints into the fusion weights. By minimizing "constraint deviation + prediction error", the weights after route fusion are dynamically corrected, and the final high-precision fusion result is output.

[0038] Therefore, based on different state triggering states, the weight of constraints is dynamically adjusted (e.g., phase constraints are emphasized in abrupt states, and trend constraints are emphasized in stable states). Simultaneously, the constraint results can be fed back to the state determination unit to optimize the accuracy of state determination in the next moment. The three paths are not simply "formulas," but rather a complete channel of "feature input, formula enhancement, state adaptation, and output": the formula is the core tool for "feature enhancement, noise suppression, and coefficient adjustment" within the path, while the path is the carrier for "frequency band feature decoupling and state adaptation." The combination of the two achieves "precise transmission of features from different frequency bands," perfectly connecting to the aforementioned state triggering mechanism, while providing a clear verification object for subsequent consistency constraints, ensuring the entire fusion logic is closed-loop and feasible.

[0039] Preferably, in this embodiment, the fusion rule of the cross-band gating fusion module is as follows: The low-frequency intrinsic mode components constrain the overall amplitude level of the high-frequency intrinsic mode components to prevent disordered high-frequency fluctuations from deviating from the trend benchmark; the phase and peak values ​​of the mid-frequency intrinsic mode components guide the timing and amplitude of the peaks in the high-frequency intrinsic mode components, ensuring that high-frequency fluctuations conform to periodic patterns; and the abnormal abrupt changes in the high-frequency intrinsic mode components are used to reverse-correct the trend and periodic fitting deviations between the mid-frequency and low-frequency intrinsic mode components, improving the response accuracy under sudden loads. (Even with low-frequency trends constraining high-frequency fluctuations, mid-frequency periodic characteristics guide high-frequency peak-valley changes, and high-frequency disturbance information is used to correct mid- and low-frequency characteristics, a structurally stable and information-rich comprehensive load sequence is generated, providing high-quality input for subsequent predictions.)

[0040] Preferably, the cross-band gating fusion module adopts a multi-band feature collaborative interaction architecture, with the noise-reduced high-frequency fluctuation IMF, mid-frequency periodic IMF, low-frequency trend IMF and RES residual trend component as parallel inputs; it includes a gating weighting unit, a contribution adaptive learning unit, a frequency band collaborative constraint unit and a feature aggregation output unit.

[0041] Specifically, in this embodiment, the residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components, and low-frequency intrinsic mode components are input together into a pre-built and trained cross-frequency band gated fusion module. This module does not simply perform linear superposition or fixed-weighted summation of the components; instead, it adaptively learns the mutual constraints between the intrinsic mode components of different frequency bands to achieve collaborative modeling and dynamic fusion of multi-frequency band information. Adaptive learning refers to the module's inclusion of a trainable contribution adaptive learning unit. This unit takes multi-dimensional information such as the load's temporal context, real-time fluctuation intensity, operating period type, and periodic characteristic phase as input. Through a fully connected layer and gradient descent backpropagation mechanism, it automatically learns and dynamically updates the contribution weights of each frequency band component in the fusion process at each moment. These weights are not pre-set fixed values ​​but are adjusted in real time according to changes in load conditions, allowing the fusion process to flexibly adapt to the changing characteristics of the power load under different time periods and operating conditions. Through this dynamic fusion method, the cross-band gating fusion module can effectively fuse the intrinsic mode components of different frequency bands into a single comprehensive load sequence. This sequence remains consistent with the original load sequence in the time dimension, but in terms of numerical characteristics, it simultaneously includes high-frequency fluctuation details, mid-frequency periodicity, low-frequency trend trends, and long-term evolution information of residual components, thus achieving efficient integration of multi-scale features. Specifically, the gating mechanism dynamically adjusts the fusion weights based on the contribution of each frequency band feature to load changes. This allows low-frequency trend information to constrain the overall level, mid-frequency periodicity information to adjust the temporal structure, and high-frequency fluctuation information to characterize short-term changes, thereby achieving effective fusion of multi-band information without independent frequency band prediction.

[0042] To further reflect the physical synergy between information from different frequency bands, this invention establishes explicit fusion rules in the cross-band gating fusion module, making the fusion process interpretable and engineering-reasonable. First, the overall amplitude level of the high-frequency intrinsic mode components is constrained by the low-frequency intrinsic mode components, i.e., low-frequency constrains high-frequency: low-frequency trend information is introduced through gating weights to constrain the overall level of high-frequency characteristics. Since the low-frequency components reflect the basic trend and long-term level of the load, high-frequency fluctuations should not deviate from this benchmark for a long time. Therefore, the module uses a gating weighting unit to dynamically scale and constrain the high-frequency components using the amplitude range of the low-frequency components, preventing disordered high-frequency fluctuations from exceeding a reasonable range, thereby improving the stability of the fusion sequence. Second, the timing and amplitude of the high-frequency intrinsic mode component peaks are guided by the phase and peak values ​​of the mid-frequency intrinsic mode components, i.e., mid-frequency guides high-frequency fluctuations: the temporal position and amplitude of high-frequency peaks are dynamically guided using the periodic characteristics of the mid-frequency. The intermediate frequency (IF) component typically corresponds to regular electricity consumption behavior on a daily or weekly cycle, and the timing of its peaks and troughs is somewhat predictable. High-frequency spikes often overlap these periodic peaks. Through a frequency band collaborative constraint unit, the module uses the phase information of the IF component to calibrate the timing of high-frequency spikes and adjusts the reasonable range of high-frequency spikes by referencing the peak amplitude of the IF component. This ensures that the fused sequence retains necessary abrupt change characteristics while conforming to basic periodicity. Third, the abnormal abrupt change information of the high-frequency intrinsic mode component is used to reverse-correct the trend and periodic fitting deviation of the IF and low-frequency intrinsic mode components; that is, high-frequency feedback corrects low-frequency: high-frequency abnormal fluctuation information is fed back to the low-frequency characteristics to improve the model's response to sudden load changes. When the load experiences a sudden jump or drop, this abnormal information is first reflected in the high-frequency component. The module uses a reverse feedback mechanism to transmit the high-frequency abrupt change information to the IF and low-frequency branches, dynamically compensating for the periodic fitting deviation of the IF and the trend tracking error of the low-frequency, thereby improving the model's response accuracy under sudden events. The three fusion rules described above are executed in parallel and cooperate with each other within the module, forming a two-way interactive constraint mechanism. This makes the fusion process no longer a one-way information compression, but a deep collaboration between multiple frequency bands. Subsequently, gated fusion calculation is performed: the features of each frequency band are weighted and fused to generate a unified comprehensive load feature sequence. Finally, the fusion result is output: the fused features are input into the fully connected layer or the sequence decoding module, and the comprehensive load sequence for multi-timescale prediction is output.

[0043] In a preferred embodiment, the aforementioned cross-band gating fusion module adopts a multi-band feature collaborative interaction architecture, which specifically includes four components: a gating weighting unit, a contribution adaptive learning unit, a frequency band collaborative constraint unit, and a feature aggregation output unit. The contribution adaptive learning unit, as described above, is responsible for learning the dynamic fusion weights of each frequency band component in real time. The frequency band collaborative constraint unit is responsible for executing the three fusion rules mentioned above: low-frequency constrains high-frequency, mid-frequency guides high-frequency, and high-frequency reverses low-frequency correction. The gating weighting unit performs time-step gating activation and weighting calculations on the four input components based on the learned dynamic weights, effectively suppressing redundant information and noise interference. The feature aggregation output unit then concatenates and nonlinearly maps the gating weighted multi-band features to ultimately generate a comprehensive load sequence that is completely aligned with the time dimension of the original load sequence, has complete information, and suppresses noise. Through this architecture, the cross-band gated fusion module realizes refined collaborative processing of multi-band information. Compared with the traditional fixed weight superposition or independent prediction and fusion method, the fusion method proposed in this invention can dynamically adapt to load changes and significantly improve the comprehensive expression ability of the fusion sequence of trend, periodic and mutation information, thereby providing more stable and accurate input features for subsequent multi-timescale prediction.

[0044] Specifically, this module incorporates a trainable adaptive contribution learning unit that dynamically learns the contribution weights of each frequency band component to load changes in real time (based on load time series, fluctuation state, and periodic characteristics). These weights are adaptively adjusted with time series and load state rather than being fixedly assigned. A gating mechanism constrains high-frequency fluctuations with low-frequency trends, guides high-frequency peaks and valleys with mid-frequency cycles, and corrects mid- and low-frequency characteristics with high-frequency disturbances. This integrates multi-frequency band features into a comprehensive load sequence, preserving key information while suppressing noise redundancy, providing stable and accurate input for prediction. In other words, the adaptive contribution learning unit takes load time series context, real-time fluctuation intensity, runtime type, periodic characteristic phase, and system state vector as multi-dimensional inputs. Through a trainable fully connected layer and gradient descent backpropagation mechanism, it automatically learns and iteratively updates the real-time contribution weights of the four components (high-frequency, mid-frequency, low-frequency, and RES) at each time step. These weights change dynamically with load state rather than being fixed values.

[0045] The gated weighting unit performs time-step gated activation and weighting calculation on the four types of input components based on the learned dynamic weights to suppress redundant information and noise interference. The feature aggregation output unit splices and nonlinearly maps the gated weighted multi-band features to generate a comprehensive load sequence that is consistent with the time dimension of the original load sequence, time-aligned, information-complete, and noise-suppressed, providing high-quality and stable input for subsequent multi-timescale prediction.

[0046] S3. Input the single comprehensive load sequence into the pre-trained multi-timescale prediction model for prediction processing to generate the final load prediction sequence. Specifically, step S3 further includes: inputting a single composite load sequence into a pre-trained multi-timescale prediction model, where the TCN unit in the multi-timescale prediction model extracts the local temporal dependence and short-term fluctuation features of the single composite load sequence through dilated causal convolution to generate local temporal features. The ABiLSTM unit in the multi-timescale prediction model is used to predict local time series features. In this model, an attention mechanism is combined to complete bidirectional deep time series modeling, capturing the long-term dependence and global trend of local time series features, and obtaining a preliminary prediction sequence. Based on the preliminary prediction sequence, a feature matrix is ​​constructed. The attention Seq2Seq unit in the multi-timescale prediction model is used to adaptively learn the feature matrix, learn the contribution weight of the intrinsic mode components of each frequency band in the feature matrix, and perform information fusion to obtain the fused prediction sequence. The residual components are used to perform trend correction on the fused prediction sequence to obtain the final load prediction sequence.

[0047] In this embodiment, a multi-timescale prediction model is constructed to perform load prediction for daily, weekly, and monthly prediction tasks respectively. The multi-timescale prediction model can adopt a deep learning time series model, including but not limited to TCN, ABiLSTM, Transformer, GRU or a combination thereof, to achieve collaborative modeling of short-term fluctuations and medium-to-long-term trends, thereby improving the prediction accuracy at different prediction timescales.

[0048] Specifically, a single composite load sequence is input into the TCN unit in a multi-timescale prediction model. The TCN unit is an in-sequence convolutional network, whose core lies in its dilated causal convolution architecture. Unlike ordinary convolution, dilated causal convolution uses an exponentially growing dilation factor to allow the convolutional kernel to cover a longer historical interval in a jump-like manner along the time dimension, while strictly ensuring causality. That is, the output at the current moment depends only on the inputs at the past and present moments, without introducing future information. The TCN unit uses this dilated causal convolution to process the input single composite load sequence layer by layer, effectively extracting local temporal dependencies and short-term fluctuation features from the sequence. Specifically, each layer of dilated causal convolution, while preserving the temporal order, captures local patterns at different scales through an ever-expanding receptive field, such as load ramp-ups, sudden drops, or spikes over several hours. After multi-layer stacking and residual connections, the TCN unit finally generates local temporal features containing rich local structural information. This feature not only preserves the short-term abrupt change details in the original composite load sequence, but also achieves a unified representation of multi-scale local patterns through the mechanism of dilated convolution. Compared to the gradient vanishing problem that traditional recurrent neural networks are prone to in long sequences, TCN units, with their parallel computing and stable gradient backpropagation characteristics, can extract local temporal features more efficiently and stably, laying a solid foundation for subsequent deep temporal modeling.

[0049] Subsequently, the ABiLSTM unit in the multi-timescale prediction model further processes the local temporal features output by the aforementioned TCN unit for prediction. The ABiLSTM unit, short for Bidirectional Long Short-Term Memory Network based on an attention mechanism, is composed of a bidirectional LSTM layer and an additive attention mechanism. Specifically, the bidirectional LSTM layer includes two processing directions: forward LSTM and backward LSTM. The forward LSTM traverses the local temporal feature sequence from front to back, capturing the influence of historical information on the current moment; the backward LSTM traverses from back to front, capturing the inverse dependence of future information on the current moment. By concatenating the hidden states from both directions, the ABiLSTM unit can simultaneously utilize past and future contextual information, thereby gaining a more complete understanding of the evolution of the load sequence. Based on this, the ABiLSTM unit further incorporates an attention mechanism to complete bidirectional deep temporal modeling. The attention mechanism calculates the correlation score between the hidden state at each time step and the current prediction target, generates a set of normalized attention weights, and then uses these weights to sum the hidden states at each time step, obtaining a context vector focusing on key information. This mechanism enables the model to automatically focus on historical moments that contribute significantly to the current prediction, such as key time windows before and after load peaks, while ignoring information less relevant to the prediction target. Through the aforementioned bidirectional deep time series modeling and attention weighting, the ABiLSTM unit effectively captures long-term dependencies and global trends in local time series features, generating a preliminary prediction sequence. This preliminary prediction sequence has already well integrated short-term fluctuations, periodic patterns, and long-term evolutionary directions, providing a foundation for subsequent refined fusion.

[0050] Building upon this, the present invention further optimizes the prediction results using an attention-based Seq2Seq unit in a multi-timescale prediction model. Specifically, the implementers construct a feature matrix based on the aforementioned preliminary prediction sequence. Each row of this feature matrix corresponds to a time step, and each column corresponds to the prediction contribution of the intrinsic mode components (IMCs) of different frequency bands. The attention-based Seq2Seq unit adopts a sequence-to-sequence architecture, where the encoder maps the feature matrix to a fixed-dimensional context vector, and the decoder generates the fused prediction sequence step by step based on this context vector. In this process, the attention-based Seq2Seq unit adaptively learns the feature matrix through a built-in attention mechanism, that is, it automatically learns the contribution weights of the IMCs of each frequency band in the feature matrix. Unlike the traditional Seq2Seq model, which treats all input information equally, the introduction of the attention mechanism allows the model to dynamically focus on the frequency band information in the feature matrix that is most relevant to the current prediction time in each decoding step. For example, during a rapid load increase phase, the model automatically increases the weights corresponding to high-frequency components; during a stable change phase, it enhances the contributions of mid-frequency and low-frequency components. In this way, the attention-based Seq2Seq unit achieves adaptive weighting and information fusion of information from each frequency band, resulting in a fused prediction sequence. This fusion process effectively overcomes the shortcomings of fixed-weight fusion methods in terms of adaptability under different operating conditions, enabling the model to dynamically adjust the fusion strategy under different load conditions, thereby improving the flexibility and accuracy of prediction.

[0051] Finally, this invention uses residual components to perform trend correction on the above-mentioned fused prediction sequence to obtain the final load prediction sequence. The residual components represent the slowest long-term evolution trend in the original load sequence, such as overall level changes influenced by economic development, seasonal changes, or macroeconomic policies. Although the fused prediction sequence already contains information from low-frequency intrinsic mode components, the residual components, as independently extracted monotonic trend terms during signal decomposition, exhibit smoother changes and are less affected by mid-to-high frequency components. Therefore, adding or weighting the fused prediction sequence and residual components step-by-step can effectively correct the deviation of the fused prediction sequence in the ultra-long-term trend, ensuring that the final output load prediction sequence retains multi-scale information on short-term fluctuations, medium-term cycles, and long-term trends, while numerically approximating the true distribution of the original load. Through this trend correction step, the final load prediction sequence has higher practical value in engineering, especially in long-term load prediction scenarios spanning seasons and years, significantly improving the stability and reliability of the prediction results.

[0052] S4. The quantile regression method is used to estimate the confidence interval of the final load forecast sequence, and the predicted values ​​and prediction confidence intervals (such as 90% confidence interval) of multiple quantiles (such as 5%, 50%, 95%) are calculated to provide a basis for power dispatch and risk assessment. Specifically, step S4 further includes: using quantile regression to optimize the quantile loss function to perform confidence interval estimation on the final load forecast sequence, calculating the predicted values ​​of multiple quantiles, and constructing the prediction confidence intervals for each future time point based on these predicted values.

[0053] In this embodiment, quantile regression is used to optimize the quantile loss function and perform confidence interval estimation on the final load forecast sequence. Unlike traditional mean regression or least squares regression, quantile regression is not limited to fitting the conditional expected value, but can estimate the conditional distribution of the dependent variable at different quantiles. In the technical solution of this invention, the specific implementation method is as follows: during the training or prediction phase of the multi-timescale prediction model, a quantile loss function is introduced as one of the optimization objectives. This loss function is calculated based on the asymmetric weights of the deviation between the predicted value and the true value at different quantiles, enabling the model to learn multiple conditional quantiles of the output load value under given input conditions. For example, implementers can set three typical quantile levels, namely 5%, 50%, and 95%. Among them, the 50% quantile corresponds to the predicted median and can be used as the output of the point prediction value; the 5% quantile and 95% quantile correspond to the lower and upper boundaries of the predicted distribution, respectively. By optimizing the quantile loss function, the model can simultaneously output multiple quantile predictions. That is, for each future time point to be predicted, it generates the corresponding 5% quantile prediction, 50% quantile prediction, and 95% quantile prediction.

[0054] After calculating the predicted values ​​for multiple quantiles, confidence intervals for future time points are constructed based on these values. Specifically, for each predicted time point, the 5th quantile predicted value is used as the lower confidence limit, and the 95th quantile predicted value is used as the upper confidence limit. The interval between these two values ​​constitutes the 90% confidence interval for that time point. This confidence interval means that, given the current input information and historical data distribution of the model, there is a 90% probability that the future actual load value will fall within this interval. Compared to a single point predicted value, this confidence interval intuitively reflects the range of uncertainty in the prediction result: when the load change is stable and the model confidence is high, the confidence interval is narrower; when the load fluctuates drastically or is in an operating condition rarely seen by the model, the confidence interval will widen accordingly. Implementation personnel can choose different combinations of quantile levels according to actual business needs, such as using the 1% and 99th quantiles to construct a 98% confidence interval, or using the 10% and 90th quantiles to construct an 80% confidence interval, flexibly adjusting the strictness of uncertainty quantification.

[0055] Through the aforementioned quantile regression and confidence interval construction methods, this step provides complete quantitative information on the uncertainty of power load forecasting results. This process has several beneficial effects. First, it significantly improves the reliability of the forecast results. Traditional point forecasting methods only output a single value, which dispatchers cannot determine the reliability of. The introduction of confidence intervals gives the forecast results probabilistic meaning, facilitating decision-makers to make reasonable judgments under different risk preferences. Second, it provides a quantitative basis for power system risk assessment. When formulating reserve capacity plans or combining generating units, dispatchers can refer to the upper and lower boundaries of the confidence interval to assess the possibility and potential impact of load exceeding the forecast range, thereby formulating more robust operating plans. Third, the quantile regression method does not require the assumption that the forecast error follows a specific distribution (such as a normal distribution). Instead, it directly learns the conditional distribution characteristics at different quantiles from the data. Therefore, it has a natural adaptability to skewed distributions, heavy-tailed distributions, and heteroscedasticity commonly found in actual power load data, and can more realistically reflect the actual form of forecast uncertainty. Fourth, this method is seamlessly integrated with the cross-band gating fusion and multi-timescale prediction module mentioned above in this invention. It can output rich probability prediction information with only a small increase in computational overhead, which significantly improves the engineering practical value of the overall technical solution.

[0056] S5 performs post-processing and prediction performance evaluation on the predicted values ​​and prediction confidence intervals of multiple quantiles, and obtains a visualization report.

[0057] Specifically, step S5 further includes: performing inverse normalization on the predicted values ​​and prediction confidence intervals of multiple quantiles to restore their dimensions to the dimensions of historical power load data; The predicted performance is evaluated using preset indicators to assess the recovered predicted values ​​and prediction confidence intervals, generating evaluation results (i.e., simultaneously using quantile regression to generate prediction confidence intervals to quantify prediction uncertainty). This results in a visualized report containing historical power load data, predictions for multiple quantiles, prediction confidence intervals, and the evaluation results. The evaluation indicators include root mean square error (RMSE), mean absolute error (MAE), mean square error (MSE), and coefficient of determination (R²). 2 ).

[0058] In this embodiment, the predicted values ​​and confidence intervals of multiple quantiles are first denormalized. Recalling the data preprocessing stage, the original historical power load data underwent normalization preprocessing before being input into the model. This typically employs maximum-minimum normalization or Z-score standardization to map all values ​​to a uniform interval, such as 0 to 1, or to make them conform to a standard normal distribution. Correspondingly, the predicted values ​​output by the model, as well as the predicted values ​​of each quantile and the upper and lower boundaries of the confidence intervals calculated based on quantile regression, are all under the same normalized dimension, making them unsuitable for direct use in actual power dispatch decisions. Therefore, denormalization is an essential subsequent step. Based on the normalization parameters recorded in step S1, such as minimum and maximum values ​​or mean and standard deviation, the implementers perform an inverse transformation on the prediction results, restoring their dimensions to the original dimensions of the historical power load data, for example, to megawatts or kilowatt-hours. After denormalization, the predicted values ​​and confidence intervals regain clear physical meaning, enabling direct comparison with the original load data and engineering applications. For example, the 50th percentile forecast for a certain point in time is 1500 MW, and the 90% confidence interval is 1450 MW to 1550 MW. These values ​​can be directly used for power generation planning or reserve capacity allocation.

[0059] After completing the denormalization process, the predicted performance of the recovered predicted values ​​and prediction confidence intervals is further evaluated using preset indicators, generating evaluation results. Specifically, implementers select a set of evaluation indicators that can comprehensively reflect the prediction accuracy and reliability. For point prediction performance, indicators such as root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAS%), and coefficient of determination are typically used. RMSE is sensitive to large deviations and can reflect the overall magnitude of prediction errors; MAE intuitively describes the average deviation between predicted and actual values; MAS is presented as a relative percentage, facilitating comparisons between load data of different magnitudes; and the coefficient of determination measures the model's explanatory power for load data variability. For probabilistic prediction performance, i.e., the quality of the confidence interval, implementers can additionally evaluate indicators such as interval coverage and average interval width. Interval coverage refers to the actual proportion of actual load values ​​falling within the prediction confidence interval, which should theoretically be close to the set confidence level; the average interval width reflects the refinement of prediction uncertainty, and a narrower interval implies higher prediction information content while ensuring coverage. Through systematic calculation of the aforementioned preset indicators, implementers can comprehensively evaluate the overall performance of the proposed method in terms of both point prediction accuracy and probabilistic prediction reliability, generating a quantitative evaluation result. This evaluation result not only verifies the effectiveness of the model but also serves as a basis for model iterative optimization and hyperparameter adjustment.

[0060] Finally, a visualization report is generated, containing historical power load data, forecasts at multiple quantiles, forecast confidence intervals, and evaluation results. This report is presented using a combination of graphs and tables. Graphically, implementers can simultaneously plot historical power load curves, point forecast curves, and confidence interval bands on a single time-series graph. Historical data is represented by one color, point forecasts by another, and confidence intervals are shown as semi-transparent shaded areas around the forecast curves, visually illustrating the range of forecast uncertainty. Forecasts at multiple quantiles, such as the 5%, 50%, and 95% quantiles, can be plotted on the same coordinate system using different line types or colors. The table summarizes the specific values ​​of each evaluation indicator, such as the root mean square error (RMSE) in megawatts, the mean absolute percentage error (MAS) in percent, and the actual interval coverage rate in percent. This visualization report facilitates rapid diagnostics of model performance by R&D personnel and allows power dispatchers to intuitively understand the potential range of future load changes and their reliability, enabling efficient use of forecast results in daily operations and emergency decision-making.

[0061] In summary, this method first performs missing value processing, outlier correction, and normalization preprocessing on historical power load data. Then, it employs the Complete Ensemble Empirical Mode Decomposition (CEEMDAN) algorithm to adaptively decompose the original non-stationary load sequence into several intrinsic mode components (IMFs) and residual components (RES). Based on the spectral characteristics and energy distribution of each IMF component, it is divided into high-frequency, mid-frequency, and low-frequency component groups to construct a multi-scale load subsequence.

[0062] Building upon this foundation, a cross-frequency gating fusion (CF-GF) mechanism is introduced to collaboratively model high, medium, and low frequency components as well as residual components. Through adaptive learning of the interrelationships between different frequency bands by the gating control unit, low-frequency trends constrain high-frequency fluctuations, medium-frequency periodicity guides load change structure, and high-frequency disturbance information is used to correct medium- and low-frequency forecasts. This achieves effective fusion of multi-frequency band information, generating a stable and fully informative comprehensive load sequence. Subsequently, the fused load sequence is input into a multi-timescale forecasting model, constructing daily, weekly, and monthly load forecasting models to achieve accurate forecasts across different forecast time ranges. Furthermore, a quantile regression method is introduced to quantify forecast uncertainty and output forecast confidence intervals. Experimental results demonstrate that this invention significantly improves the accuracy, stability, and practical value of power load forecasting.

[0063] Compared with existing technologies, this method has the following advantages: 1. It adopts CEEMDAN decomposition combined with optimization algorithms, which can adaptively process non-stationary load signals and improve the adaptability and accuracy of prediction models for each component. 2. It adopts a parallel prediction channel and TCN-ABiLSTM hybrid architecture to achieve targeted modeling and deep feature extraction for different frequency components. 3. It introduces an attention mechanism into the ABiLSTM module, enabling the model to automatically focus on key time-series information, enhancing prediction accuracy and interpretability. 4. It achieves dynamic fusion of multi-branch prediction results through CF-GF cross-band gating fusion and Seq2Seq model, and combines residual components for trend correction, improving fusion effect and stability. 5. It introduces quantile regression method to generate prediction confidence intervals, effectively quantifying uncertainty and improving the credibility of prediction results. 6. It can adopt a modular design, has good scalability and maintainability, and can be widely used in the field of power system load forecasting.

[0064] Please see Figure 2 A second embodiment of the present invention provides a multi-time-scale accurate power load forecasting device, which includes: The decomposition unit 101 is used to acquire historical power load data, perform data preprocessing and signal decomposition on the historical power load data, and obtain residual components, high-frequency intrinsic mode component groups, mid-frequency intrinsic mode component groups and low-frequency intrinsic mode component groups. The fusion unit 102 is used to input the residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components and low-frequency intrinsic mode components into a preset cross-band gated fusion module. The module adaptively learns the mutual constraint relationship between the intrinsic mode components of different frequency bands to achieve collaborative modeling and dynamic fusion of multi-frequency band information, and fuses the intrinsic mode components of different frequency bands into a single comprehensive load sequence. The prediction unit 103 is used to input a single comprehensive load sequence into a pre-trained multi-timescale prediction model for prediction processing, and generate the final load prediction sequence. The estimation unit 104 is used to perform confidence interval estimation on the final load forecast sequence using the quantile regression method, and calculate the predicted values ​​and prediction confidence intervals of multiple quantiles. Evaluation unit 105 is used to perform post-processing and prediction performance evaluation on the predicted values ​​and prediction confidence intervals of multiple quantiles, and to obtain a visualization report.

[0065] A third embodiment of the present invention provides a multi-timescale power load accurate prediction device, which includes a memory and a processor. The memory stores a computer program, which can be executed by the processor to implement the multi-timescale power load accurate prediction method as described in any of the above embodiments.

[0066] The fourth embodiment of the present invention provides a computer-readable storage medium, characterized in that it stores a computer program, which can be executed by the processor of the device in which the computer-readable storage medium is located, to realize the multi-time-scale power load accurate prediction method as described in any of the above.

[0067] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for accurate prediction of power load across multiple time scales, characterized in that, include: Historical power load data is acquired, and the historical power load data is preprocessed and decomposed to obtain residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components, and low-frequency intrinsic mode components. The residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components, and low-frequency intrinsic mode components are input into a preset cross-band gated fusion module. The mutual constraint relationship between the intrinsic mode components of different frequency bands is adaptively learned to achieve collaborative modeling and dynamic fusion of multi-band information, and the intrinsic mode components of different frequency bands are fused into a single comprehensive load sequence. The single comprehensive load sequence is input into a pre-trained multi-timescale prediction model for prediction processing to generate the final load prediction sequence. The quantile regression method is used to estimate the confidence interval of the final load forecast sequence, and the predicted values ​​and prediction confidence intervals of multiple quantiles are calculated. Post-processing and prediction performance evaluation are performed on the predicted values ​​and prediction confidence intervals of multiple quantiles to obtain a visualization report.

2. The method for accurate prediction of multi-timescale power load according to claim 1, characterized in that, Historical power load data is acquired, and data preprocessing and signal decomposition are performed on the historical power load data to obtain residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components, and low-frequency intrinsic mode components, specifically: Historical power load data is acquired and preprocessed. The preprocessing includes missing value repair, outlier removal, and normalization. The preprocessed historical power load data is adaptively decomposed using the complete set empirical mode decomposition algorithm to obtain intrinsic mode component groups and residual components. The intrinsic mode components in the intrinsic mode component group are arranged from high to low frequency. The intrinsic mode components are input into the preset IMF-GAN adversarial noise reduction module for noise reduction, resulting in a clean multi-band load subsequence. Based on the spectral characteristics and energy distribution of the multi-band load subsequence, the multi-band load subsequence is divided into high-frequency intrinsic mode component group, mid-frequency intrinsic mode component group and low-frequency intrinsic mode component group.

3. The method for accurate prediction of multi-timescale power load according to claim 1, characterized in that, The residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components, and low-frequency intrinsic mode components are input into a preset cross-band gated fusion module. The module adaptively learns the mutual constraints between the intrinsic mode components of different frequency bands to achieve collaborative modeling and dynamic fusion of multi-band information, fusing the intrinsic mode components of different frequency bands into a single comprehensive load sequence. Specifically: The load operation status determination unit is used to determine the high-frequency intrinsic mode component group, the mid-frequency intrinsic mode component group, and the low-frequency intrinsic mode component group, classifying the current load status into a stationary state, a volatile dynamic state, an abrupt change state, and a periodically dominant state. A three-component Gaussian mixture model is used to fit the probability distribution of the fluctuation intensity sequence to construct a mixture distribution. Three Gaussian distributions were obtained. , For a steady state, It is a periodic dominant state. The state is abruptly changed, where V is the wave intensity and k is the Gaussian component index. The mixing weight coefficients for the k-th Gaussian component are... It follows a Gaussian distribution. Let be the mean of the k-th Gaussian component. Let be the variance of the k-th Gaussian component; Based on the current load status, corresponding paths are assigned, and feature enhancement is performed on the corresponding frequency band within each path. Cross-path collaborative fusion is carried out at the exit of all paths to form a preliminary fusion result. Specifically, when the load status is in a stable state, low-frequency components enter the trend path, mid-frequency components weakly participate in the periodic path, and high-frequency components are suppressed. When the load status is in a periodic dominant state, mid-frequency components enter the periodic path, low-frequency components enter the trend path, and high-frequency components enter the abrupt path. When the load status is in an abrupt state, high-frequency components enter the abrupt path, mid-frequency components enter the periodic path, and low-frequency components enter the trend path. Based on the residual components, the preliminary fusion results are subjected to cross-frequency consistency constraint fusion processing to generate a single comprehensive load sequence. The cross-frequency consistency constraint fusion processing includes trend consistency constraint, phase consistency constraint, energy conservation constraint and constraint optimization fusion. The fusion rules of the cross-band gating fusion module are as follows: the low-frequency intrinsic mode components constrain the overall amplitude level of the high-frequency intrinsic mode components; the phase and peak value of the mid-frequency intrinsic mode components guide the occurrence time and amplitude of the peaks of the high-frequency intrinsic mode components; and the abnormal change information of the high-frequency intrinsic mode components is used to reversely correct the trend and period fitting deviation of the mid-frequency intrinsic mode components and the low-frequency intrinsic mode components.

4. The method for accurate prediction of multi-timescale power load according to claim 3, characterized in that, The cross-band gating fusion module adopts a multi-band feature collaborative interaction architecture, which includes a gating weighting unit, a contribution adaptive learning unit, a frequency band collaborative constraint unit, and a feature aggregation output unit.

5. The method for accurate prediction of multi-timescale power load according to claim 1, characterized in that, The single composite load sequence is input into a pre-trained multi-timescale prediction model for prediction processing to generate the final load prediction sequence, specifically: The single composite load sequence is input into the pre-trained multi-timescale prediction model. The TCN unit in the multi-timescale prediction model extracts the local temporal dependence and short-term fluctuation features of the single composite load sequence through dilated causal convolution, and generates local temporal features. The ABiLSTM unit in the multi-timescale prediction model is used to predict local time series features. In this model, an attention mechanism is combined to complete bidirectional deep time series modeling, capturing the long-term dependence and global trend of local time series features, and obtaining a preliminary prediction sequence. Based on the preliminary prediction sequence, a feature matrix is ​​constructed. The attention Seq2Seq unit in the multi-timescale prediction model is used to adaptively learn the feature matrix, learn the contribution weight of the intrinsic mode components of each frequency band in the feature matrix, and perform information fusion to obtain the fused prediction sequence. The residual components are used to perform trend correction on the fused prediction sequence to obtain the final load prediction sequence.

6. The method for accurate prediction of multi-timescale power load according to claim 1, characterized in that, The quantile regression method is used to estimate the confidence interval of the final load forecast sequence, and the predicted values ​​and prediction confidence intervals of multiple quantiles are calculated. Specifically, the quantile regression method is used to optimize the quantile loss function to estimate the confidence interval of the final load forecast sequence, the predicted values ​​of multiple quantiles are calculated, and the prediction confidence intervals of future time points are constructed based on these predicted values.

7. The method for accurate prediction of multi-timescale power load according to claim 1, characterized in that, Post-processing and prediction performance evaluation are performed on the predicted values ​​and prediction confidence intervals of multiple quantiles to obtain a visualization report, as follows: The predicted values ​​and prediction confidence intervals of multiple quantiles are denormalized to restore their dimensions to the dimensions of historical power load data. The predicted performance is evaluated using preset indicators to assess the recovered predicted values ​​and prediction confidence intervals, generating evaluation results and a visualized report that includes historical power load data, predictions for multiple quantiles, prediction confidence intervals, and evaluation results.

8. A multi-timescale power load accurate prediction device, characterized in that, include: The decomposition unit is used to acquire historical power load data, perform data preprocessing and signal decomposition on the historical power load data, and obtain residual components, high-frequency intrinsic mode component groups, mid-frequency intrinsic mode component groups and low-frequency intrinsic mode component groups. The fusion unit is used to perform fusion and collaborative processing on residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components and low-frequency intrinsic mode components using cross-band gated fusion modules to generate a single integrated load sequence; The prediction unit is used to input the residual components, high-frequency intrinsic mode components, mid-frequency intrinsic mode components and low-frequency intrinsic mode components into a preset cross-band gated fusion module. The module adaptively learns the mutual constraint relationship between the intrinsic mode components of different frequency bands to achieve collaborative modeling and dynamic fusion of multi-frequency band information, and fuses the intrinsic mode components of different frequency bands into a single comprehensive load sequence. The estimation unit is used to perform confidence interval estimation on the final load forecast sequence using the quantile regression method, and calculates the predicted values ​​and prediction confidence intervals for multiple quantiles. The evaluation unit is used to perform post-processing and prediction performance evaluation on the predicted values ​​and prediction confidence intervals of multiple quantiles, and to obtain a visual report.

9. A multi-timescale power load accurate prediction device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be executed by the processor to implement the multi-timescale power load accurate prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that can be executed by a processor of the device in which the computer-readable storage medium is located, to implement the multi-timescale power load accurate prediction method as described in any one of claims 1 to 7.

Images