Deep learning-based multi-time scale-oriented load prediction method

Abstract

The invention is suitable for the technical field of load prediction, and provides a deep learning-based multi-time scale-oriented load prediction method, which comprises the following steps: acquiring a multi-scale feature coding field of historical power load data, the field comprising a time scale identifier and a dynamic complexity parameter; if the time scale identifier corresponds to the time scale and the dynamic complexity parameter meets the threshold range, selecting a multi-mode space-time fusion model, a lightweight time sequence decomposition model or a space-time and frequency spectrum joint model or an integrated model to carry out load prediction; calculating an error between each prediction result and a true value; and adjusting weight parameters of the multi-modal space-time fusion model, the lightweight time sequence decomposition model, the space-time and spectrum joint model or the weighted fusion strategy through errors. Through multi-scale coding, dynamic complexity driving model selection and adaptive fusion, the prediction stability in a complex power scene is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and in particular to a deep learning-based load forecasting method for multiple time scales. Background Technology

[0002] Power load forecasting is one of the core technologies for optimal energy system scheduling. Currently, power load forecasting mainly relies on statistical analysis using methods that assume linearity, such as ARIMA and exponential smoothing; short-term forecasts are achieved by automatically extracting time-series features through deep learning models; and spatiotemporal forecasts are made by modeling spatial correlations between regions using graph neural networks. However, single-time-scale models struggle to adapt to the interaction between minute-level high-frequency fluctuations and daily trends, leading to long-term forecasting interference from high-frequency noise. Furthermore, spatiotemporal feature fusion often depends on static graph structures, ignoring the dynamic spatial correlation of load propagation and failing to incorporate frequency domain information, resulting in errors in the prediction of sudden events. These problems severely limit the prediction accuracy and robustness in complex power scenarios. Summary of the Invention

[0003] This invention provides a deep learning-based load forecasting method for multiple time scales, which addresses the issues of accuracy and robustness in power load forecasting under complex power scenarios.

[0004] This invention provides a deep learning-based load forecasting method for multiple time scales, comprising:

[0005] The multi-scale feature encoding field of historical power load data is obtained. The multi-scale feature encoding field includes a time scale identifier and a dynamic complexity parameter. The time scale identifier corresponds to the minute scale, hour scale and day scale respectively. The dynamic complexity parameter is calculated by time-frequency wavelet entropy and spatial correlation measure.

[0006] If the time scale identifier corresponds to an hourly or daily scale, and the dynamic complexity parameter meets the first threshold range, then the multimodal spatiotemporal fusion model is selected.

[0007] If the time scale identifier corresponds to a minute scale and the dynamic complexity parameter meets the second threshold range, then the lightweight time series decomposition model is selected.

[0008] If the time scale identifier corresponds to a minute scale, an hour scale, and a day scale, and the dynamic complexity parameter meets the third threshold range, then the spatiotemporal and spectral joint model is selected.

[0009] If the dynamic complexity parameter is in the critical interval of adjacent threshold ranges, a weighted fusion strategy is used to integrate and predict the model outputs corresponding to adjacent threshold ranges.

[0010] Calculate the error between the first prediction result output by the multimodal spatiotemporal fusion model, the second prediction result output by the lightweight temporal decomposition model, the third prediction result output by the spatiotemporal and spectral joint model, or the fourth prediction result output by the weighted fusion strategy and the true value;

[0011] When the calculated error exceeds a preset threshold, the weight parameters of the multimodal spatiotemporal fusion model, the lightweight temporal decomposition model, the spatiotemporal and spectral joint model, or the weighted fusion strategy are adjusted.

[0012] Furthermore, the dynamic complexity parameter is calculated from the time-frequency wavelet entropy and spatial correlation metric, including:

[0013] If the time scale identifier is an hourly or daily scale, the time-frequency wavelet entropy is calculated on the target area load time series data in the historical power load data to obtain a first time-frequency domain complexity index; wherein, when the first time-frequency domain complexity index meets a preset condition, the spatial correlation measure of time series data in different areas is calculated, and a dynamic complexity parameter is generated based on the first time-frequency domain complexity index and the spatial correlation measure; when the first time-frequency domain complexity index does not meet the preset condition, the first time-frequency domain complexity index is directly used as the dynamic complexity parameter.

[0014] If the time scale identifier is a minute scale, then the time-frequency wavelet entropy is calculated for the target area load time series data in the historical power load data to obtain the second time-frequency domain complexity index; and the spatial correlation measure of time series data in different areas is calculated simultaneously, and a dynamic complexity parameter is generated based on the second time-frequency domain complexity index and the spatial correlation measure.

[0015] Furthermore, the preset condition is that the first time-frequency domain complexity index exceeds a preset independent threshold. The preset independent threshold is determined by the statistical distribution of the time-frequency domain complexity index of historical data. Specifically, it is calculated based on the quantile values ​​of all hourly or daily time-frequency domain complexity indices in historical data, and the quantile values ​​are used as the independent threshold.

[0016] Furthermore, if the time scale identifier is an hourly or daily scale, the time-frequency wavelet entropy is calculated on the target area load time-series data in the historical power load data to obtain a first time-frequency domain complexity index, including:

[0017] The target area load time series data is decomposed into multiple frequency band sub-signals through wavelet transform. The frequency band sub-signals include low-frequency components reflecting long-term trends, mid-to-low-frequency components reflecting daily cycle characteristics, and mid-frequency components reflecting hourly cycle characteristics.

[0018] Calculate the energy proportion of each frequency band sub-signal and generate time-frequency wavelet entropy based on the energy proportion distribution; use the time-frequency wavelet entropy as the first time-frequency domain complexity index.

[0019] Furthermore, when the first time-frequency domain complexity index meets a preset condition, the spatial correlation measure of time-series data in different regions is calculated, and a dynamic complexity parameter is generated based on the first time-frequency domain complexity index and the spatial correlation measure, including:

[0020] Obtain time-series data from at least two adjacent regions associated with the target region;

[0021] Based on time series data within a preset time window, calculate the mutual information or dynamic time warping distance between the target area and each adjacent area.

[0022] The maximum mutual information value or the minimum dynamic time warp distance is used as a measure of spatial correlation.

[0023] The first time-frequency domain complexity index and spatial correlation metric are normalized and then weighted and summed according to preset weight coefficients to generate dynamic complexity parameters.

[0024] Furthermore, if the time scale identifier is a minute scale, then the time-frequency wavelet entropy is calculated on the target area load time-series data in the historical power load data to obtain the second time-frequency domain complexity index, including:

[0025] The target area load time series data is decomposed into high-frequency sub-signals and ultra-high-frequency sub-signals through wavelet transform. The high-frequency sub-signals reflect minute-level periodic fluctuations, and the ultra-high-frequency sub-signals reflect transient random fluctuations.

[0026] The time-frequency wavelet entropy is calculated based on the energy distribution of the high-frequency and ultra-high-frequency sub-signals to obtain the second time-frequency domain complexity index; wherein, the frequency band division of the high-frequency and ultra-high-frequency sub-signals is dynamically adjusted according to the real-time sampling frequency.

[0027] Furthermore, if the time scale identifier corresponds to an hourly or daily scale, and the dynamic complexity parameter meets the first threshold range, then selecting a multimodal spatiotemporal fusion model includes:

[0028] Load time-series data of the target area and load time-series data of adjacent areas are input into a multimodal spatiotemporal fusion model, which makes predictions through the following steps:

[0029] Spatial features of multi-region load time series data are extracted using convolutional neural networks;

[0030] Long Short-Term Memory (LSTM) networks are used to capture the temporal dependencies of load in the target region;

[0031] Spatial and temporal features are fused using an attention mechanism to generate prediction results;

[0032] The dynamic complexity parameter corresponding to the first threshold range indicates that the load sequence has a medium to high degree of spatiotemporal coupling characteristics.

[0033] Furthermore, if the time scale identifier corresponds to a minute scale and the dynamic complexity parameter satisfies the second threshold range, then selecting a lightweight time series decomposition model includes:

[0034] Time series decomposition is performed on minute-level load time series data of the target area to obtain trend, periodic and residual terms;

[0035] Each decomposed component is input into a sub-network of a lightweight temporal decomposition model for prediction. The sub-network is a temporal convolutional network or a lightweight recurrent neural network.

[0036] The outputs of each subnetwork are reconstructed to generate the final prediction result;

[0037] The dynamic complexity parameter corresponding to the second threshold range indicates that the load sequence has low complexity and is dominated by local transient fluctuations.

[0038] Furthermore, if the time scale identifier corresponds to a minute scale, an hour scale, and a day scale, and the dynamic complexity parameter satisfies the third threshold range, then the spatiotemporal and spectral joint model is selected, including:

[0039] The following processing is performed on the multi-scale load time series data of the target area for synchronization:

[0040] Frequency domain features are extracted using wavelet transform to construct a spectral energy distribution matrix;

[0041] Using graph neural networks to model the spatial topology of multi-regional loads;

[0042] The temporal, frequency, and spatial features are input into the spatiotemporal attention module for joint encoding to generate prediction results.

[0043] The dynamic complexity parameter corresponding to the third threshold range characterizes the load sequence as having cross-scale, non-stationary, and strong spatiotemporal spectral coupling characteristics.

[0044] Furthermore, if the dynamic complexity parameter is within a critical interval of adjacent threshold ranges, a weighted fusion strategy is used to integrate and predict the model outputs corresponding to adjacent threshold ranges, including:

[0045] The weight coefficients of each model are determined based on the distance ratio between the dynamic complexity parameter and the adjacent threshold.

[0046] The prediction results of two models that meet adjacent threshold ranges are weighted and summed to generate an integrated prediction result;

[0047] When the dynamic complexity parameter approaches the critical range between the first and third threshold ranges, the weights of the spatiotemporal and spectral joint model are preferentially increased.

[0048] When the dynamic complexity parameter is close to the critical range between the first threshold range and the second threshold range, the weights of the multimodal spatiotemporal fusion model and the lightweight temporal decomposition model are preferentially increased.

[0049] As can be seen from the above technical solutions, the present invention has the following advantages:

[0050] This invention quantifies the time-frequency non-stationarity and spatial correlation of load at minute, hourly, and daily scales through joint encoding of time scale identifiers and dynamic complexity parameters. Based on adaptive model selection using dynamic complexity parameters and threshold ranges, it improves computational efficiency and prediction accuracy for different time scales and complexity scenarios through lightweight time-series decomposition, multimodal spatiotemporal fusion, or spatiotemporal spectrum joint models. Dynamic weight allocation and priority rules smoothly transition multi-model prediction results within threshold critical intervals, avoiding the predictive value jumps caused by traditional hard switching. Real-time monitoring of model prediction errors and dynamic adjustment of weight parameters enhance the adaptability to sudden load pattern changes and reduce the need for manual intervention. Through multi-scale encoding, dynamic complexity-driven model selection, and adaptive fusion mechanisms, the invention effectively improves prediction stability and generalization in complex power scenarios. Attached Figure Description

[0051] Figure 1 This is a schematic flowchart of an embodiment of a deep learning-based load forecasting method for multiple time scales in this invention.

[0052] Figure 2 This is a schematic diagram of a partial embodiment of a deep learning-based load forecasting method for multiple time scales in this invention. Detailed Implementation

[0053] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0054] Example 1

[0055] The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The following section will describe the deep learning-based multi-timescale load forecasting method of this application from the perspective of system implementation. Please refer to... Figure 1 The method provided in this application includes the following steps:

[0056] S11. Obtain the multi-scale feature coding field of historical power load data. The multi-scale feature coding field includes time scale identifier and dynamic complexity parameter. The time scale identifier corresponds to minute scale, hour scale and day scale respectively. The dynamic complexity parameter is calculated by time-frequency wavelet entropy and spatial correlation measure.

[0057] In this embodiment, historical power load data is converted into standardized feature fields containing time scale identifiers and dynamic complexity parameters through multi-scale feature encoding. First, according to the set sampling interval, the data is divided into three scales: minute, hour, and day, and a corresponding time scale identifier, such as "M", "H", and "D" codes, is assigned to each data segment.

[0058] The calculation of the dynamic complexity parameter is divided into two parts: For time-frequency wavelet entropy, wavelet transform is used to decompose the target region load sequence into low-frequency, mid-frequency, and high-frequency sub-signals. The energy proportion of each sub-signal is calculated, and entropy values ​​are generated to quantify the non-stationarity of the sequence. For spatial correlation measurement, based on multi-region load data within a sliding window, the correlation strength between the target region and adjacent regions is calculated using mutual information or dynamic time warping algorithms. The maximum mutual information value or the minimum dynamic time warping distance is used as the spatial correlation index. After normalizing the time-frequency wavelet entropy and the spatial correlation index, they are linearly fused according to preset weights to generate the dynamic complexity parameter, providing a quantitative basis for subsequent model selection.

[0059] S12. If the time scale identifier corresponds to an hourly or daily scale, and the dynamic complexity parameter meets the first threshold range, then select the multimodal spatiotemporal fusion model.

[0060] Load time series data of the target area and load time series data of adjacent areas are input into a multimodal spatiotemporal fusion model, which then makes predictions through the following steps:

[0061] 1. Use convolutional neural networks to extract spatial features from multi-regional load time series data;

[0062] 2. Use Long Short-Term Memory (LSTM) networks to capture the temporal dependencies of the load in the target region;

[0063] 3. Spatial and temporal features are fused using an attention mechanism to generate prediction results;

[0064] Among them, the dynamic complexity parameter corresponding to the first threshold range characterizes the load sequence as having a medium to high degree of spatiotemporal coupling.

[0065] Specifically, when the dynamic complexity parameter is within the first threshold range, it indicates that there is a significant spatiotemporal correlation between the load sequence of the target area and the load of adjacent areas, such as inter-regional power consumption linkage at the diurnal scale. It is necessary to jointly model spatiotemporal features to capture such dependencies. The multimodal spatiotemporal fusion model extracts multi-regional spatial features through convolutional neural networks (CNN), models the temporal dependence of the target area through long short-term memory networks (LSTM), and then dynamically fuses spatiotemporal information through an attention mechanism.

[0066] Spatial Feature Extraction (CNN): Input load time series data of the target region and adjacent regions to construct a spatiotemporal matrix (time dimension × spatial dimension); use multi-layer convolutional kernels (such as 3×3) to extract local spatial patterns (such as load fluctuation propagation paths between regions); compress redundant information through pooling operations to generate spatial feature vectors.

[0067] Temporal Dependency Modeling (LSTM): Input the target region load time series data into the LSTM network, and use gating mechanisms (forget gate, input gate, output gate) to capture long-term and short-term time series dependencies; output a time series feature vector to characterize the load trend, periodicity and residual components.

[0068] Attention mechanism fusion: Calculate the correlation score between spatial features and temporal features to generate attention weights; fuse the two types of features according to the weights to generate the final prediction result.

[0069] The first threshold range corresponds to the numerical range of the dynamic complexity parameter. Its lower limit is a preset independent threshold, such as the 75th quantile mentioned above, and its upper limit is the maximum value of the dynamic complexity parameter at the hourly / day scale in historical data. For example, if the independent threshold is 0.85 and the historical maximum value is 1.2, then the first threshold range is 0.85 to 1.2. A dynamic complexity parameter ≥ 0.85 indicates that the parameter generated jointly by time-frequency wavelet entropy and spatial correlation indicates that the load sequence has a medium to high spatiotemporal coupling strength; a parameter < 0.85 indicates that the parameter generated only by time-frequency wavelet entropy indicates weak spatiotemporal coupling, and there is no need to enable the multimodal spatiotemporal fusion model. If a new region is added to the system or the load pattern changes abruptly, the quantile threshold and range range are recalculated to ensure the adaptability of the model selection.

[0070] S13. If the time scale identifier corresponds to the minute scale and the dynamic complexity parameter meets the second threshold range, then select the lightweight time series decomposition model.

[0071] 1. Perform time-series decomposition on minute-level load time-series data of the target area to obtain trend, periodic, and residual terms;

[0072] 2. Input the decomposed components into the sub-networks of the lightweight temporal decomposition model for prediction. The sub-networks are either temporal convolutional networks or lightweight recurrent neural networks.

[0073] 3. Reconstruct the outputs of each subnetwork to generate the final prediction results;

[0074] Among them, the dynamic complexity parameter corresponding to the second threshold range indicates that the load sequence has low complexity and is dominated by local transient fluctuations.

[0075] Specifically, when the dynamic complexity parameter is within the second threshold range, it indicates that the time-frequency energy distribution of the minute-level load sequence in the target region is concentrated and the spatial correlation is weak. This type of scenario is dominated by transient fluctuations, and the long-term trend and periodic characteristics are not significant. The lightweight time series decomposition model decomposes the original sequence into trend terms, periodic terms, and residual terms, which are independently predicted by a lightweight quantum network. This can effectively isolate noise interference and reduce computational overhead.

[0076] The temporal decomposition employs adaptive local weighted regression or wavelet packet decomposition to break down minute-level load time-series data into: a trend term reflecting long-term, slow changes at the minute level; a periodic term capturing periodic fluctuations at the minute level; and a residual term representing transient random noise. Sub-network prediction includes: applying causal dilated convolution to the trend term to capture long-range dependencies; employing simplified gating units for the periodic term to reduce the number of parameters while modeling short-term patterns; directly filtering and discarding residuals if their energy is below a preset threshold; otherwise, inputting them into a lightweight temporal convolutional network for prediction. Finally, the prediction results of each sub-network are superimposed in reverse decomposition to generate the final prediction sequence.

[0077] The second threshold range corresponds to the numerical range of the dynamic complexity parameter. Its upper limit is the preset independent threshold, and its lower limit is the minimum value of the dynamic complexity parameter at the minute scale in historical data. For example, if the independent threshold is 0.85 and the historical minimum value is 0.3, then the second threshold range is [0.3, 0.85]. A dynamic complexity parameter < 0.85 indicates that the sequence complexity is low and spatial correlation is not activated, with the load mainly consisting of local transient fluctuations. A parameter ≥ 0.85 triggers a multimodal spatiotemporal fusion model or a spatiotemporal spectrum joint model. The minimum value and independent threshold of the minute-scale dynamic complexity parameter are periodically recalculated to ensure that the threshold range adapts to system changes.

[0078] S14. If the time scale identifier corresponds to the minute scale, hour scale, and day scale, and the dynamic complexity parameter meets the third threshold range, then select the spatiotemporal and spectrum joint model.

[0079] The following processing is performed on the multi-scale load time series data of the target area for synchronization:

[0080] 1. Extract frequency domain features using wavelet transform and construct a spectral energy distribution matrix;

[0081] 2. Use graph neural networks to model the spatial topology of multi-region loads;

[0082] 3. Input the time domain, frequency domain, and spatial features into the spatiotemporal attention module for joint encoding to generate prediction results;

[0083] Among them, the dynamic complexity parameter corresponding to the third threshold range characterizes the load sequence as having cross-scale, non-stationary, and strong spatiotemporal spectral coupling characteristics.

[0084] Specifically, when the dynamic complexity parameter is within the third threshold range, it indicates that the target region's load sequence simultaneously exhibits minute-level transient fluctuations, hourly-level periodic fluctuations, and daily-level trend changes, with a highly non-stationary time-frequency energy distribution and strong spatial correlation with adjacent regions. The spatiotemporal and spectral joint model extracts multi-scale frequency domain features through wavelet transform, models the dynamic spatial topology using graph neural networks, and fuses time-frequency-spatial multi-dimensional features using a spatiotemporal attention mechanism to achieve holographic modeling of cross-scale complex patterns.

[0085] First, wavelet packet transform is applied to the minute, hour, and day-scale load time-series data of the target area, decomposing them into multiple frequency band sub-signals, such as ultra-high frequency, high frequency, medium frequency, and low frequency. The energy proportion of each frequency band is calculated, constructing a cross-scale spectral energy distribution matrix, where rows represent time scales, columns represent frequency bands, and values ​​represent energy intensity, quantifying non-stationary characteristics. Second, a dynamic graph structure is constructed based on the power grid topology (such as regional connectivity and transmission capacity), with nodes representing regional load data and edge weights dynamically updated through real-time mutual information or impedance values. A graph attention network is used to aggregate features of neighboring nodes, capturing the spatial dependence of load propagation. Finally, the time-domain (original load sequence), frequency-domain (energy matrix), and spatial (graph attention network output) features are input into the spatiotemporal attention module; cross-modal feature correlations are calculated through a multi-head attention mechanism, weights are dynamically allocated, and the results are fused to generate predictions.

[0086] The third threshold range corresponds to the numerical range of the dynamic complexity parameter, with its lower limit being the upper limit of the first threshold range and its upper limit being the maximum value of the dynamic complexity parameter in historical data. For example, the third threshold range is 1.2 to 2.0. A dynamic complexity parameter ≥ 1.2 indicates that the load sequence has strong cross-scale coupling characteristics, meaning it is generated jointly by time-frequency wavelet entropy, spatial correlation, and frequency domain energy. A parameter < 1.2 indicates that only a multimodal spatiotemporal fusion model or a lightweight decomposition model needs to be used.

[0087] S15. If the dynamic complexity parameter is in the critical interval of adjacent threshold ranges, a weighted fusion strategy is used to integrate and predict the model outputs corresponding to adjacent threshold ranges.

[0088] 1. Determine the weight coefficients of each model based on the distance ratio between the dynamic complexity parameter and adjacent thresholds;

[0089] 2. Weighted summation of the prediction results of two models that meet adjacent threshold ranges to generate an integrated prediction result;

[0090] 3. When the dynamic complexity parameter is close to the critical range between the first and third threshold ranges, prioritize increasing the weight of the spatiotemporal and spectral joint model;

[0091] 4. When the dynamic complexity parameter is close to the critical range between the first threshold range and the second threshold range, prioritize increasing the weights of the multimodal spatiotemporal fusion model and the lightweight temporal decomposition model.

[0092] Specifically, the critical interval is when the dynamic complexity parameter lies between two adjacent threshold ranges, such as the boundary between the first threshold range of 0.85, 1.2 and the third threshold range of 1.2, 2.0, at 1.15, 1.25. Model switching may cause a sudden change in the prediction result. Let the boundary values ​​of the adjacent threshold ranges be T1 and T2, and the dynamic complexity parameter be D, then the formula for calculating the weight coefficient is: Where w1 is the weight of the model closer to the high threshold range, and w2 is the weight of the model closer to the low threshold range.

[0093] Dynamic allocation of weighting coefficients: If the dynamic complexity parameter D = 1.18, and it falls within the critical interval between the first threshold range (0.85, 1.2) and the third threshold range (1.2, 2.0), then: Correction rule: To avoid excessive weight bias towards low-threshold models, set a minimum weight threshold (e.g., 0.1). When w1 < 0, force w1 = 0.1 and w2 = 0.9.

[0094] The prediction results Y1 and Y2 from adjacent models are fused according to their weights: Y 融合 = w1Y1 + w2Y2; When D approaches the critical region between the first and third thresholds, the weights of the spatiotemporal and spectral joint model are forcibly increased because it is more suitable for cross-scale coupling scenarios. When D approaches the critical region between the first and second thresholds, the weights of the multimodal spatiotemporal fusion model and the lightweight temporal decomposition model are proportionally allocated.

[0095] S16. Calculate the error between the first prediction result output by the multimodal spatiotemporal fusion model, the second prediction result output by the lightweight temporal decomposition model, the third prediction result output by the spatiotemporal and spectral joint model, or the fourth prediction result output by the weighted fusion strategy and the true value.

[0096] This step quantifies the model's prediction accuracy at corresponding time scales and complexity scenarios by calculating the error between each model's prediction and the actual value, providing a data foundation for error feedback and weight adjustment. Specifically, it dynamically evaluates the stability and reliability of model performance by statistically analyzing the error distribution through a sliding window.

[0097] S17. When the calculated error exceeds the preset threshold, adjust the weight parameters of the multimodal spatiotemporal fusion model, the lightweight temporal decomposition model, the spatiotemporal and spectral joint model, or the weighted fusion strategy.

[0098] This step is based on a model weight adaptive adjustment mechanism triggered by error exceeding the limit. It uses gradient descent or Bayesian optimization to dynamically update the model weight parameters, such as the fused weights or sub-network parameters mentioned above, to minimize the prediction error. If the error continues to exceed the limit, the model is forced to be retrained or switched to an alternative strategy to ensure system robustness.

[0099] Example 2

[0100] Please see Figure 2 The dynamic complexity parameter is calculated from the time-frequency wavelet entropy and spatial correlation metric, including the following steps:

[0101] S111. If the time scale identifier is hourly or daily, calculate the time-frequency wavelet entropy of the target area load time series data in the historical power load data to obtain the first time-frequency domain complexity index.

[0102] Specifically, when the first time-frequency domain complexity index meets the preset conditions, the spatial correlation measure of time series data in different regions is calculated, and a dynamic complexity parameter is generated based on the first time-frequency domain complexity index and the spatial correlation measure; when the first time-frequency domain complexity index does not meet the preset conditions, the first time-frequency domain complexity index is directly used as the dynamic complexity parameter.

[0103] In this embodiment, the preset condition is that the first time-frequency domain complexity index exceeds a preset independent threshold. The preset independent threshold is determined by the statistical distribution of the time-frequency domain complexity index of historical data. Specifically, the quantile value is calculated based on the time-frequency domain complexity index of all hourly or daily scales in the historical data, and the quantile value is used as the independent threshold.

[0104] Specifically, the preset conditions are based on the following: when the time-frequency complexity of the load sequence is high, it indicates significant non-stationarity or multi-frequency band coupling characteristics, requiring further introduction of spatial correlation metrics to improve prediction accuracy. Quantile calculation: Based on the time-frequency domain complexity indices of all hourly or daily scales in historical data, their distribution is statistically analyzed according to time windows, and specific quantiles are selected as independent thresholds. For example, if the 75th quantile of historical wavelet entropy values ​​is 0.85, then the independent threshold is set to 0.85. If the system load pattern changes significantly, such as seasonal changes causing a shift in frequency domain energy distribution, the quantiles are recalculated to ensure the thresholds are adapted to the current data characteristics.

[0105] In this embodiment, the first time-frequency domain complexity index is obtained through the following steps:

[0106] 1. The target area load time series data is decomposed into multiple frequency band sub-signals through wavelet transform. The frequency band sub-signals include low-frequency components reflecting long-term trends, mid-to-low frequency components reflecting daily cycle characteristics, and mid-frequency components reflecting hourly cycle characteristics.

[0107] 2. Calculate the energy proportion of each frequency band sub-signal and generate time-frequency wavelet entropy based on the energy proportion distribution; use the time-frequency wavelet entropy as the first time-frequency domain complexity index.

[0108] Specifically, the wavelet transform here uses the Daubechies wavelet basis, where the low-frequency component (0.001-0.1Hz) reflects long-term trends, such as monthly load changes; the mid-to-low-frequency component (0.1-0.5Hz) reflects daily periodic characteristics, such as morning and evening peak electricity consumption; and the mid-frequency component (0.5-2Hz) reflects hourly fluctuations, such as intermittent start-up and shutdown of industrial equipment. The number of decomposition levels is adaptively selected based on the data length.

[0109] Calculate the energy value E of each frequency band sub-signal i : Where x i (k) represents the k-th data point of the i-th frequency band sub-signal, and N is the data length; calculate the energy proportion p of each frequency band. i : Where M is the total number of frequency bands, E j Let H be the energy value of the j-th frequency band sub-signal; generate the time-frequency wavelet entropy H: The larger the entropy value, the more dispersed the time-frequency energy distribution of the load sequence and the higher the complexity.

[0110] In this embodiment, when the first time-frequency domain complexity index meets the preset conditions, the spatial correlation measure of time series data in different regions is calculated, and dynamic complexity parameters are generated based on the first time-frequency domain complexity index and the spatial correlation measure, including the following steps:

[0111] 1. Obtain time-series data from at least two adjacent regions associated with the target region;

[0112] Load time-series data from at least two adjacent regions directly connected to the target region's power grid topology are extracted from the power database to ensure the physical correlation of the spatial correlation analysis.

[0113] 2. Based on time-series data within a preset time window, calculate the mutual information or dynamic time warping distance between the target area and each adjacent area;

[0114] The preset time window is a 24-hour sliding window. Based on the load sequence within this sliding window, the KL divergence between the joint probability distribution and the marginal distribution of the target area and each adjacent area is calculated to quantify the degree of information sharing.

[0115] Align the time axes of the target region and the load sequences of adjacent regions, calculate the minimum path cumulative distance, and resolve the problem of inconsistent sequence lengths.

[0116] 3. Use the maximum mutual information value or the minimum dynamic time-warped distance as a measure of spatial correlation;

[0117] The maximum mutual information value reflects strong correlation, while the minimum dynamic time warp distance reflects morphological similarity.

[0118] 4. Normalize the first time-frequency domain complexity index and the spatial correlation measure, and sum them according to the preset weight coefficients to generate dynamic complexity parameters.

[0119] The time-frequency complexity index and spatial correlation metric are normalized to the [0,1] interval respectively. The preset weights are set according to the actual scenario. The time-frequency complexity and spatial correlation are balanced by weighted fusion to accurately represent the dynamic characteristics of the load.

[0120] S112. If the time scale identifier is minute scale, calculate the time-frequency wavelet entropy of the target area load time series data in the historical power load data to obtain the second time-frequency domain complexity index; and simultaneously calculate the spatial correlation measure of time series data in different areas, and generate dynamic complexity parameters based on the second time-frequency domain complexity index and spatial correlation measure.

[0121] In this embodiment, the second time-frequency domain complexity index is obtained, including the following:

[0122] 1. The target area load time series data is decomposed into high-frequency sub-signals and ultra-high-frequency sub-signals through wavelet transform. The high-frequency sub-signals reflect minute-level periodic fluctuations, while the ultra-high-frequency sub-signals reflect transient random fluctuations.

[0123] Here, the wavelet transform is an adaptive wavelet packet transform. The high-frequency sub-signal is 2-10Hz, reflecting minute-level periodic fluctuations, such as frequent elevator starts and stops; the ultra-high-frequency sub-signal is 10-30Hz, reflecting transient random fluctuations, such as instantaneous overload caused by lightning strikes. If the sampling frequency is 1 minute / sample, the high-frequency sub-signal corresponds to a period of 2-5 minutes; if the sampling frequency is 10 seconds / sample, the high-frequency sub-signal extends to a period of 5-30 seconds.

[0124] 2. The time-frequency wavelet entropy is calculated based on the energy distribution of the high-frequency and ultra-high-frequency sub-signals to obtain the second time-frequency domain complexity index; wherein, the frequency band division of the high-frequency and ultra-high-frequency sub-signals is dynamically adjusted according to the real-time sampling frequency.

[0125] The energy proportions of high-frequency and ultra-high-frequency sub-signals are calculated separately to generate time-frequency wavelet entropy, using the same method as in step S111, thus obtaining the second time-frequency domain complexity index. Multi-region mutual information or dynamic time warping distance calculations are performed in parallel to ensure the real-time requirements of minute-level data. The final dynamic complexity parameter generation rule is consistent with step S111. This step, with its high-frequency / ultra-high-frequency decomposition, accurately captures minute-level transient characteristics, and its parallel computation ensures real-time performance, adapting to the needs of power grid emergency event prediction.

[0126] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods.

[0127] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based load forecasting method for multiple time scales, characterized in that, include: The multi-scale feature encoding field of historical power load data is obtained. The multi-scale feature encoding field includes a time scale identifier and a dynamic complexity parameter. The time scale identifier corresponds to the minute scale, hour scale and day scale respectively. The dynamic complexity parameter is calculated by time-frequency wavelet entropy and spatial correlation measure. If the time scale identifier corresponds to an hourly or daily scale, and the dynamic complexity parameter meets the first threshold range, then the multimodal spatiotemporal fusion model is selected. If the time scale identifier corresponds to a minute scale and the dynamic complexity parameter meets the second threshold range, then the lightweight time series decomposition model is selected. If the time scale identifier corresponds to a minute scale, an hour scale, and a day scale, and the dynamic complexity parameter meets the third threshold range, then the spatiotemporal and spectral joint model is selected. If the dynamic complexity parameter is in the critical interval of adjacent threshold ranges, a weighted fusion strategy is used to integrate and predict the model outputs corresponding to adjacent threshold ranges. Calculate the error between the first prediction result output by the multimodal spatiotemporal fusion model, the second prediction result output by the lightweight temporal decomposition model, the third prediction result output by the spatiotemporal and spectral joint model, or the fourth prediction result output by the weighted fusion strategy and the true value; When the calculated error exceeds a preset threshold, the weight parameters of the multimodal spatiotemporal fusion model, the lightweight temporal decomposition model, the spatiotemporal and spectral joint model, or the weighted fusion strategy are adjusted.

2. The deep learning-based load forecasting method for multiple time scales according to claim 1, characterized in that, The dynamic complexity parameter is calculated from the time-frequency wavelet entropy and spatial correlation metric, including: If the time scale identifier is an hourly or daily scale, the time-frequency wavelet entropy is calculated on the target area load time series data in the historical power load data to obtain a first time-frequency domain complexity index; wherein, when the first time-frequency domain complexity index meets a preset condition, the spatial correlation measure of time series data in different areas is calculated, and a dynamic complexity parameter is generated based on the first time-frequency domain complexity index and the spatial correlation measure; when the first time-frequency domain complexity index does not meet the preset condition, the first time-frequency domain complexity index is directly used as the dynamic complexity parameter. If the time scale identifier is a minute scale, then the time-frequency wavelet entropy is calculated for the target area load time series data in the historical power load data to obtain the second time-frequency domain complexity index; and the spatial correlation measure of time series data in different areas is calculated simultaneously, and a dynamic complexity parameter is generated based on the second time-frequency domain complexity index and the spatial correlation measure.

3. The deep learning-based load forecasting method for multiple time scales according to claim 2, characterized in that, The preset condition is that the first time-frequency domain complexity index exceeds a preset independent threshold. The preset independent threshold is determined by the statistical distribution of the time-frequency domain complexity index of historical data. Specifically, it is calculated based on the quantile values ​​of all hourly or daily time-frequency domain complexity indices in historical data, and the quantile values ​​are used as the independent threshold.

4. The deep learning-based load forecasting method for multiple time scales according to claim 2, characterized in that, If the time scale identifier is an hourly or daily scale, the time-frequency wavelet entropy is calculated on the target area load time-series data in the historical power load data to obtain a first time-frequency domain complexity index, including: The target area load time series data is decomposed into multiple frequency band sub-signals through wavelet transform. The frequency band sub-signals include low-frequency components reflecting long-term trends, mid-to-low-frequency components reflecting daily cycle characteristics, and mid-frequency components reflecting hourly cycle characteristics. Calculate the energy proportion of each frequency band sub-signal and generate time-frequency wavelet entropy based on the energy proportion distribution; use the time-frequency wavelet entropy as the first time-frequency domain complexity index.

5. The deep learning-based load forecasting method for multiple time scales according to claim 2, characterized in that, When the first time-frequency domain complexity index meets the preset conditions, the spatial correlation measure of time series data in different regions is calculated, and dynamic complexity parameters are generated based on the first time-frequency domain complexity index and the spatial correlation measure, including: Obtain time-series data from at least two adjacent regions associated with the target region; Based on time series data within a preset time window, calculate the mutual information or dynamic time warping distance between the target area and each adjacent area. The maximum mutual information value or the minimum dynamic time warp distance is used as a measure of spatial correlation. The first time-frequency domain complexity index and spatial correlation metric are normalized and then weighted and summed according to preset weight coefficients to generate dynamic complexity parameters.

6. The deep learning-based load forecasting method for multiple time scales according to claim 5, characterized in that, If the time scale identifier is a minute scale, then the time-frequency wavelet entropy is calculated on the target area load time-series data in the historical power load data to obtain the second time-frequency domain complexity index, including: The target area load time series data is decomposed into high-frequency sub-signals and ultra-high-frequency sub-signals through wavelet transform. The high-frequency sub-signals reflect minute-level periodic fluctuations, and the ultra-high-frequency sub-signals reflect transient random fluctuations. The time-frequency wavelet entropy is calculated based on the energy distribution of the high-frequency and ultra-high-frequency sub-signals to obtain the second time-frequency domain complexity index; wherein, the frequency band division of the high-frequency and ultra-high-frequency sub-signals is dynamically adjusted according to the real-time sampling frequency.

7. The deep learning-based load forecasting method for multiple time scales according to claim 1, characterized in that, If the time scale identifier corresponds to an hourly or daily scale, and the dynamic complexity parameter meets the first threshold range, then a multimodal spatiotemporal fusion model is selected, including: Load time-series data of the target area and load time-series data of adjacent areas are input into a multimodal spatiotemporal fusion model, which makes predictions through the following steps: Spatial features of multi-region load time series data are extracted using convolutional neural networks; Long Short-Term Memory (LSTM) networks are used to capture the temporal dependencies of load in the target region; Spatial and temporal features are fused using an attention mechanism to generate prediction results; The dynamic complexity parameter corresponding to the first threshold range indicates that the load sequence has a medium to high degree of spatiotemporal coupling characteristics.

8. The deep learning-based load forecasting method for multiple time scales according to claim 1, characterized in that, If the time scale identifier corresponds to a minute scale and the dynamic complexity parameter meets the second threshold range, then a lightweight time series decomposition model is selected, including: Time series decomposition is performed on minute-level load time series data of the target area to obtain trend, periodic and residual terms; Each decomposed component is input into a sub-network of a lightweight temporal decomposition model for prediction. The sub-network is a temporal convolutional network or a lightweight recurrent neural network. The outputs of each subnetwork are reconstructed to generate the final prediction result; The dynamic complexity parameter corresponding to the second threshold range indicates that the load sequence has low complexity and is dominated by local transient fluctuations.

9. The deep learning-based load forecasting method for multiple time scales according to claim 1, characterized in that, If the time scale identifier corresponds to a minute scale, an hour scale, and a day scale, and the dynamic complexity parameter meets the third threshold range, then the spatiotemporal and spectral joint model is selected, including: The following processing is performed on the multi-scale load time series data of the target area for synchronization: Frequency domain features are extracted using wavelet transform to construct a spectral energy distribution matrix; Using graph neural networks to model the spatial topology of multi-regional loads; The temporal, frequency, and spatial features are input into the spatiotemporal attention module for joint encoding to generate prediction results. The dynamic complexity parameter corresponding to the third threshold range characterizes the load sequence as having cross-scale, non-stationary, and strong spatiotemporal spectral coupling characteristics.

10. The deep learning-based load forecasting method for multiple time scales according to claim 1, characterized in that, If the dynamic complexity parameter is within the critical interval of adjacent threshold ranges, a weighted fusion strategy is used to integrate and predict the model outputs corresponding to adjacent threshold ranges, including: The weight coefficients of each model are determined based on the distance ratio between the dynamic complexity parameter and the adjacent threshold. The prediction results of two models that meet adjacent threshold ranges are weighted and summed to generate an integrated prediction result; When the dynamic complexity parameter approaches the critical range between the first and third threshold ranges, the weights of the spatiotemporal and spectral joint model are preferentially increased. When the dynamic complexity parameter is close to the critical range between the first threshold range and the second threshold range, the weights of the multimodal spatiotemporal fusion model and the lightweight temporal decomposition model are preferentially increased.

Images