Micro-grid load prediction method based on deep learning

Abstract

The invention discloses a micro-grid load prediction method based on deep learning, and the method comprises the following steps: S1, collecting micro-grid load data, and generating a load time sequence; s2, executing load structure decomposition to generate load subsequences; s3, executing load operation scene identification, and generating a load operation scene label and a load structure subsequence; s4, inputting the load structure sub-sequence into an improved SCINet model, introducing a load structure perception guide mechanism, and outputting a load prediction sequence; s5, executing consistency constraint processing, and generating a consistency load prediction sequence; s6, generating prediction deviation data; and S7, executing adaptive updating, and outputting an updated load prediction sequence. According to the method, high-precision prediction and continuous adaptive updating of the micro-grid load time sequence under the conditions of multiple operation scenes and multiple time scales are realized, and the stability and reliability of a load prediction result are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis and prediction technology for power systems, and in particular to a microgrid load prediction method based on deep learning. Background Technology

[0002] With the continuous increase in the proportion of distributed energy integration and the increasing complexity of microgrid operation, the demand for accurate prediction and intelligent dispatch of microgrid load variation characteristics is constantly growing. Microgrid load forecasting methods have become an important research direction in the field of intelligent power system operation and energy management. Existing microgrid load forecasting technologies mostly rely on traditional statistical models or single deep learning structures to model and analyze historical load data, but they generally suffer from the following problems in practical applications:

[0003] Microgrid loads are influenced by multiple factors, including fluctuations in distributed power generation output, changes in user electricity consumption behavior, and shifts in operating scenarios. Load time series exhibit significant nonlinear, non-stationary, and multi-scale superposition characteristics. Traditional holistic modeling methods struggle to effectively characterize long-term trend changes, structural differences between short- and medium-term fluctuations and abrupt load changes, resulting in insufficient sensitivity of prediction results to load structure changes. Existing load forecasting methods typically ignore the impact of operating scenarios on load evolution patterns, failing to effectively correlate load operating states with load structural characteristics, and exhibiting limited generalization ability of prediction models across different operating scenarios. Furthermore, the lack of consistency constraints among multi-timescale prediction results easily leads to deviations between short-term and medium- to long-term predictions, affecting the stability and reliability of microgrid dispatch decisions. Therefore, it is necessary to propose a microgrid load forecasting method that integrates load structure decomposition, operating scenario identification, and multi-timescale consistency constraints to improve the accuracy and robustness of load forecasting under complex operating conditions.

[0004] Therefore, how to provide a deep learning-based method for microgrid load forecasting is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a microgrid load forecasting method based on deep learning. This invention addresses the needs of microgrid load time series modeling and forecasting by introducing an improved SCINet model with a load structure perception guidance mechanism through load time alignment and scale normalization, load structure decomposition, load operation scenario identification, and load structure and scenario association modeling. Combined with multi-time scale consistency constraints and an adaptive update process for prediction deviation, this invention completes the generation and dynamic optimization of microgrid load forecasting sequences. It has the advantages of strong load structure characterization capability, high adaptability to operation scenarios, good prediction stability, and high long-term prediction accuracy.

[0006] A microgrid load forecasting method based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect microgrid load data, perform time alignment and scale normalization processing, and generate load time series;

[0008] S2. Perform load structure decomposition based on the load time series to generate load subsequences;

[0009] S3. Based on the load time series, perform load operation scenario identification, generate load operation scenario labels, and associate the load operation scenario labels with the load subsequence to generate a load structure subsequence;

[0010] S4. Input the load structure subsequence into the improved SCINet model. The improved SCINet model introduces a load structure awareness guidance mechanism and outputs a load prediction sequence.

[0011] S5. Construct a multi-time-scale load forecast sequence based on the load forecast sequence, and perform consistency constraint processing on the multi-time-scale load forecast sequence to generate a consistent load forecast sequence.

[0012] S6. Compare the consistent load prediction sequence with the corresponding actual load sequence to generate prediction deviation data;

[0013] S7. Based on the prediction deviation data, perform adaptive updates on the improved SCINet model and output the updated load prediction sequence.

[0014] Optionally, S1 specifically includes:

[0015] The load power data generated during the operation of the microgrid is collected, and the load power data is time-aligned according to a unified sampling period to form a load data set with continuous time index;

[0016] The load data set that has completed time alignment is subjected to scale normalization processing to map the load power values ​​to a uniform value range, generating standardized load data.

[0017] Based on standardized load data that has undergone scale normalization, a load time series is constructed in chronological order.

[0018] Optionally, S2 specifically includes:

[0019] The load time series is segmented according to a preset time window, and the load change amplitude sequence and load change rate sequence are calculated based on the load power values ​​at adjacent time points.

[0020] Based on the load change amplitude sequence and load change rate sequence, load structure decomposition processing is performed on the load time series to divide the load time series into long-term change components, medium- and short-term change components and abrupt change components.

[0021] The long-term variation components are recombined in chronological order to generate a basic load subsequence; the medium- and short-term variation components are recombined in chronological order to generate a fluctuating load subsequence; and the abrupt change components are recombined in chronological order to generate an abrupt load subsequence. The load subsequences include a basic load subsequence, a fluctuating load subsequence, and an abrupt load subsequence.

[0022] Optionally, the step of performing load operation scenario identification based on load time series and generating load operation scenario labels specifically includes:

[0023] The load time series is segmented according to a preset time window, and the load power numerical sequence within the corresponding time window is extracted.

[0024] Based on the load power numerical sequence, load change amplitude index, load change rate index and load level index are calculated to construct a load operation characteristic sequence;

[0025] Based on the load operation feature sequence, the load time series is subjected to operation status discrimination processing to generate a load operation status sequence;

[0026] Based on the load operation status sequence, load operation scenario labels corresponding to the load time sequence are generated in chronological order.

[0027] Optionally, associating the load operation scenario label with the load subsequence to generate a load structure subsequence specifically involves:

[0028] A scenario tag sequence is constructed for the load operation scenario tags in chronological order, and the scenario tag sequence and the load time series use the same time index.

[0029] Based on the consistency of time index between the scene label sequence and the load time series, time slice mapping is performed on the scene label sequence to divide the scene label sequence into scene label segments with the same segment boundaries as the load time series.

[0030] Based on the consistency of the time index between the scene tag fragment and the basic load subsequence, fluctuating load subsequence, and abrupt load subsequence, tag binding processing is performed on the basic load subsequence, fluctuating load subsequence, and abrupt load subsequence respectively to generate the basic load structure subsequence, fluctuating load structure subsequence, and abrupt load structure subsequence.

[0031] The load structure subsequence is generated by combining the basic load structure subsequence, the fluctuating load structure subsequence, and the abrupt load structure subsequence.

[0032] Optionally, the improved SCINet model includes an input processing structure, a recursive sequence decomposition structure, a structure-guided adjustment structure, and a prediction output structure, specifically:

[0033] The input processing structure is used to receive the load structure sub-sequence and form an input sequence according to the time index;

[0034] The recursive sequence decomposition structure is used to perform multi-level recursive sequence decomposition on the input sequence to generate a multi-level decomposed sequence.

[0035] The structure-guided adjustment structure introduces a load structure perception guidance mechanism. Based on the load structure identifiers corresponding to the basic load subsequence, fluctuating load subsequence and abrupt load subsequence, and combined with the load operation scenario label, a decomposition guidance weight is generated. The decomposition guidance weight is then applied to the recursive sequence decomposition structure to adjust the decomposition ratio of each recursive decomposition layer.

[0036] The prediction output structure is used to perform aggregation processing on the multi-level decomposition sequence output by the recursive sequence decomposition structure in chronological order to generate a load prediction sequence.

[0037] Optionally, S4 specifically includes:

[0038] The load structure subsequences are input into the improved SCINet model in time index order to construct the input sequence;

[0039] Based on the input sequence, the corresponding load structure identification information is extracted from the basic load subsequence, fluctuating load subsequence and abrupt load subsequence in the improved SCINet model to form a load structure identification sequence corresponding to the time index;

[0040] Based on the load structure identifier sequence and load operation scenario label, the decomposition guidance weights are calculated in chronological order to construct a decomposition guidance weight sequence that is consistent with the time index of the model input sequence.

[0041] During the recursive sequence decomposition process, the decomposition guiding weight sequence is introduced into the recursive sequence decomposition calculation path, and the decomposition ratio of trend change component and fluctuation change component in each recursive decomposition layer is adjusted to generate the recursive decomposition output sequence.

[0042] The recursive decomposition output sequence obtained after the recursive sequence decomposition process is processed by sequence aggregation in chronological order to generate the load prediction sequence.

[0043] Optionally, S5 specifically includes:

[0044] Based on the load forecast sequence and the preset time scale set, the load forecast sequence is resampled according to the time aggregation rule to generate a multi-time scale load forecast sequence. The multi-time scale load forecast sequence includes the forecast sub-sequences corresponding to the preset time scale set.

[0045] Alignment processing is performed on multi-timescale load forecast sequences according to time index to construct time-scale aligned forecast sequences;

[0046] Based on the time-scale aligned prediction sequence, consistency verification is performed on the prediction sub-sequences at different time scales according to the time aggregation rules to generate a consistency deviation sequence;

[0047] Consistency constraint processing is performed on the time-scale aligned prediction sequence based on the consistency deviation sequence to generate a consistent load prediction sequence.

[0048] Optionally, S6 specifically includes:

[0049] Obtain the actual load sequence that matches the time index of the consistent load forecast sequence;

[0050] The consistent load forecast sequence and the actual load sequence are aligned according to the time index to form the forecast load sequence and the actual load sequence;

[0051] A load prediction error sequence is generated by performing difference calculation processing based on the predicted load sequence and the actual load sequence.

[0052] Prediction bias data are constructed based on the load forecasting error sequence in chronological order.

[0053] Optionally, S7 specifically includes:

[0054] Obtain forecast deviation data and construct a deviation sample sequence based on the forecast deviation data. The deviation sample sequence includes the forecast load sequence, the actual load sequence, and the load forecast error sequence.

[0055] The biased sample sequence is input into the improved SCINet model, and parameter update processing is performed based on the biased sample sequence to generate updated model parameters;

[0056] Based on the prediction deviation data, the decomposed guiding weights are updated to generate the updated decomposed guiding weights.

[0057] The updated model parameters and the updated decomposition-guided weights are written into the improved SCINet model, and the updated load prediction sequence is output.

[0058] The beneficial effects of this invention are:

[0059] (1) By performing time alignment and scale normalization on the microgrid load data, a load time series with continuous time index is constructed, which effectively reduces the interference of different sampling periods and load amplitude differences on prediction modeling and improves the stability and modelability of the load time series. In the load structure modeling stage, the load time series is divided into basic load subsequence, fluctuating load subsequence and abrupt load subsequence by load structure decomposition, which enhances the ability to distinguish long-term changes, short-term fluctuations and abrupt behavior, and avoids the problem of structural feature aliasing caused by single sequence modeling.

[0060] (2) Based on the load time series, load operation scenario identification is performed, load operation scenario labels are generated, and the load operation scenario labels are associated with the load subsequence to form a load structure subsequence. This enables the load prediction process to simultaneously perceive load structure characteristics and changes in operating status, thereby improving the adaptability of the prediction results to changes in the operating environment. In the prediction modeling stage, an improved SCINet model with a load structure perception guidance mechanism is introduced. By decomposing the guidance weights to adjust the decomposition ratio of the recursive sequence, targeted modeling of time series characteristics under different load structures and operating scenarios is achieved, thereby improving the ability of the load prediction sequence to characterize complex load patterns.

[0061] (3) In the prediction result processing stage, a multi-time-scale load prediction sequence is constructed, and consistency constraint processing is performed on the multi-time-scale load prediction sequence to effectively suppress the inconsistency deviation between prediction results at different time scales and enhance the overall coordination and stability of the load prediction sequence in short-term and medium-to-long-term prediction. At the same time, by comparing the consistent load prediction sequence with the actual load sequence, prediction deviation data is generated, and the improved SCINet model is adaptively updated based on the prediction deviation data to realize the dynamic adjustment of model parameters and decomposition guidance weights, and continuously improve the load prediction accuracy and model generalization ability. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a flowchart of a microgrid load forecasting method based on deep learning proposed in this invention;

[0064] Figure 2 This is a schematic diagram of the structure of the improved SCINet model proposed in this invention;

[0065] Figure 3 This is a data flow diagram of a microgrid load forecasting method based on deep learning proposed in this invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1-3 A microgrid load forecasting method based on deep learning includes the following steps:

[0068] S1. Collect microgrid load data, perform time alignment and scale normalization processing, and generate load time series;

[0069] S2. Perform load structure decomposition based on the load time series to generate load subsequences;

[0070] S3. Based on the load time series, perform load operation scenario identification, generate load operation scenario labels, and associate the load operation scenario labels with the load subsequence to generate a load structure subsequence;

[0071] S4. Input the load structure subsequence into the improved SCINet model. The improved SCINet model introduces a load structure awareness guidance mechanism and outputs a load prediction sequence.

[0072] S5. Construct a multi-time-scale load forecast sequence based on the load forecast sequence, and perform consistency constraint processing on the multi-time-scale load forecast sequence to generate a consistent load forecast sequence.

[0073] S6. Compare the consistent load prediction sequence with the corresponding actual load sequence to generate prediction deviation data;

[0074] S7. Based on the prediction deviation data, perform adaptive updates on the improved SCINet model and output the updated load prediction sequence.

[0075] This implementation method collects microgrid load data and performs time alignment and scale normalization processing to construct a load time series, reducing the impact of different sampling periods and dimensional differences on load modeling and improving the temporal consistency and analyzability of load data. Furthermore, by performing load structure decomposition on the load time series to generate load subsequences, the long-term variation characteristics, short- and medium-term fluctuation characteristics, and abrupt change characteristics in load changes are separated and expressed, enhancing the structural identifiability of load features. Simultaneously, by performing load operation scenario identification based on the load time series and generating load operation scenario labels, these labels are associated with the load subsequences to form a load structure subsequence, ensuring that load features and operation scenario information are consistently expressed in the time dimension, improving the adaptability of load forecasting to changes in the operating environment. Finally, the load structure subsequence is input into a load structure perception guidance mechanism. The improved SCINet model incorporates load structure and scenario information to guide the decomposition path during the recursive sequence decomposition process, enhancing the model's ability to model different load structures and thus improving the accuracy of the load forecast sequence in depicting non-stationary load changes. Simultaneously, it constructs multi-timescale load forecast sequences based on the load forecast sequences and performs consistency constraint processing to suppress the offset between forecast results at different time scales, improving the consistency and stability of forecast results across multiple time scales. Furthermore, it generates prediction deviation data by comparing the consistent load forecast sequence with the actual load sequence, quantifying the prediction error. At the same time, it performs adaptive updates on the improved SCINet model based on the prediction deviation data, continuously adjusting model parameters and decomposition guidance weights according to changes in load characteristics, thereby achieving a simultaneous improvement in microgrid load forecast accuracy and long-term operational stability.

[0076] In this embodiment, S1 specifically refers to:

[0077] Load power data continuously recorded by metering devices during microgrid operation is collected. The load power data includes active power values ​​collected at fixed time intervals. Time alignment processing is performed on the load power data according to a unified sampling period. The time alignment processing includes mapping load power data from different collection time points to a unified time axis and performing interpolation to fill in missing load power data. The interpolation method adopts linear interpolation of load power values ​​at adjacent time points. Preferably, the unified sampling period is set to 15 minutes.

[0078] After time alignment processing, the load data set is subjected to scale normalization processing. The scale normalization processing includes obtaining the maximum load power value and the minimum load power value in the load data set, and performing linear mapping on the load power value based on the calculation rule of subtracting the minimum load power value from the load power value and dividing by the difference between the maximum load power value and the minimum load power value, so as to map the load power value to a unified value range of 0 to 1 and generate standardized load data.

[0079] Based on the standardized load data that has undergone scale normalization, the standardized load data is arranged and combined in chronological order to form a load time series with continuous time index.

[0080] In this embodiment, S2 specifically refers to:

[0081] The load time series is segmented according to a preset time window, which is used to limit the time span of load change analysis. The time window contains multiple consecutive sampling points. Preferably, the length of the preset time window is set to 8 or 16 sampling points. Based on the load power values ​​of adjacent time points within the time window, the load change amplitude sequence and the load change rate sequence are calculated. The load change amplitude sequence is composed of the difference between the maximum and minimum load power values ​​within the time window, and the load change rate sequence is composed of the ratio of the difference in load power values ​​of adjacent time points to the sampling time interval.

[0082] Based on the load change amplitude sequence and load change rate sequence, load structure decomposition processing is performed on the load time series. The load structure decomposition processing includes dividing the load time series into components according to the numerical range of load change amplitude and load change rate. When the load change amplitude is lower than a first threshold and the load change rate is lower than a second threshold, the load change within the corresponding time window is determined to be a long-term change component. When the load change amplitude is between the first threshold and the third threshold and the load change rate is between the second threshold and the fourth threshold, the load change within the corresponding time window is determined to be a medium- or short-term change component. When the load change amplitude is higher than the third threshold or the load change rate is higher than the fourth threshold, the load change within the corresponding time window is determined to be a sudden change component. Preferably, the first threshold is set to 0.1 of the standardized value of the load time series, the second threshold is set to 0.05, the third threshold is set to 0.3, and the fourth threshold is set to 0.2.

[0083] Load data identified as long-term variable components are reorganized according to time index order to form a basic load subsequence reflecting the slow trend of load change; load data identified as medium- and short-term variable components are reorganized according to time index order to form a fluctuating load subsequence reflecting periodic and volatile characteristics; load data identified as abrupt change components are reorganized according to time index order to form an abrupt load subsequence reflecting the characteristics of sudden load increases or decreases; through the reorganization process, a load subsequence set including the basic load subsequence, fluctuating load subsequence, and abrupt load subsequence is constructed for load operation scenario association and load prediction modeling process.

[0084] In this embodiment, the step of performing load operation scenario identification based on load time series and generating load operation scenario labels specifically includes:

[0085] The load time series is segmented according to a preset time window. The preset time window is consistent with the time window used in the load structure decomposition step. Each time window contains multiple consecutive sampling points. The load power value sequence within the time window is extracted as the input data for operation scenario identification.

[0086] The load change amplitude index, load change rate index, and load level index are calculated based on the load power numerical sequence. The load change amplitude index is composed of the difference between the maximum and minimum load power values ​​within the time window. The load change rate index is composed of the ratio of the difference between the load power values ​​at adjacent time points to the sampling time interval. The load level index is composed of the arithmetic mean of the load power values ​​within the time window. These are combined in chronological order to form a load operation characteristic sequence.

[0087] Based on the load operation characteristic sequence, the load time series is processed for operation status discrimination. The operation status discrimination process includes comparing the load change amplitude index with a preset amplitude discrimination parameter and comparing the load change rate index with a preset rate discrimination parameter. When the load change amplitude index is lower than the amplitude stability discrimination parameter and the load change rate index is lower than the rate stability discrimination parameter, the operation status in the corresponding time window is determined to be a stable operation status. When the load change amplitude index is between the amplitude stability discrimination parameter and the amplitude fluctuation discrimination parameter and the load change rate index is between the rate stability discrimination parameter and the rate fluctuation parameter, the operation status in the corresponding time window is determined to be a fluctuating operation status. When the load change amplitude index is higher than the amplitude fluctuation discrimination parameter or the load change rate index is higher than the rate fluctuation discrimination parameter, the operation status in the corresponding time window is determined to be a sudden change operation status.

[0088] Preferably, the amplitude stability discrimination parameter is set to 0.1 of the standardized load value, the rate stability discrimination parameter is set to 0.05, the amplitude fluctuation discrimination parameter is set to 0.3, and the rate fluctuation discrimination parameter is set to 0.2.

[0089] Based on the determined operating status, a load operating status sequence is constructed according to the time index order, and the stable operating status, fluctuating operating status and sudden operating status are mapped to the corresponding load operating scenario labels, forming a load operating scenario label sequence that is consistent with the load time series in the time dimension.

[0090] In this embodiment, associating the load operation scenario label with the load subsequence to generate the load structure subsequence specifically involves:

[0091] The load operation scenario labels are arranged in the order of the time index of the load time series to form a scenario label sequence that is completely consistent with the load time series in terms of sampling period and time span. Each label element in the scenario label sequence corresponds to a time index position in the load time series.

[0092] Based on the consistency relationship between the scene label sequence and the load time series in the time index dimension, the scene label sequence is processed by time slice mapping according to the segmentation boundary used in the load structure decomposition step of the load time series, so that each scene label segment corresponds to a complete load time segment interval, and the scene label segment contains scene label information that corresponds one-to-one with all load sampling points in the time segment interval.

[0093] Based on the consistency relationship between scene tag segments and basic load subsequences in the time index dimension, scene tag information belonging to the same time segment interval in the scene tag segment is bound to the load value at the corresponding time position in the basic load subsequence to form a basic load structure subsequence, where each load value is associated with the corresponding running scene identifier.

[0094] Meanwhile, based on the consistency relationship between scene tag segments and fluctuating load subsequences in the time index dimension, scene tag information belonging to the same time segment interval in the scene tag segment is bound to the load value at the corresponding time position in the fluctuating load subsequence, forming a fluctuating load structure subsequence, so that the change of fluctuating load and the state of the running scene are expressed synchronously in the time dimension.

[0095] Furthermore, based on the consistency relationship between scene label fragments and mutation load subsequences in the time index dimension, scene label information belonging to the same time segment interval in the scene label fragments is bound to the load value at the corresponding time position in the mutation load subsequence, forming a mutation load structure subsequence, so that the mutation load change is clearly associated with the corresponding running scene state;

[0096] The basic load structure subsequence, fluctuating load structure subsequence, and abrupt load structure subsequence are combined and processed according to a unified time index order to form a load structure subsequence containing load structure type identifier, operation scenario label, and load value information.

[0097] In this embodiment, the improved SCINet model includes an input processing structure, a recursive sequence decomposition structure, a structure-guided adjustment structure, and a prediction output structure, specifically:

[0098] The input processing structure receives a load structure subsequence, which consists of a basic load structure subsequence, a fluctuating load structure subsequence, and a sudden load structure subsequence. Each load value is associated with a corresponding load structure identifier and a load operation scenario label. The input processing structure arranges and combines the load structure subsequences according to a unified sampling period and time index order to construct a multi-dimensional input sequence containing load values, load structure identifiers, and load operation scenario labels. The time index serves as the main sequence dimension, while the load structure identifier and load operation scenario label serve as auxiliary feature dimensions.

[0099] The recursive sequence decomposition structure performs a layer-by-layer recursive decomposition of the input sequence along the time index direction. In each recursive decomposition layer, the input sequence is divided into trend change components and fluctuation change components. The recursive sequence decomposition structure performs repeated decomposition of the trend change components and fluctuation change components in a multi-layer stacking manner to generate multi-layer decomposed sequences with different time scale characteristics.

[0100] The structure-guided adjustment structure introduces a load structure perception guidance mechanism during the recursive sequence decomposition process. This mechanism generates decomposition guidance weights based on the load structure identifier and load operation scenario label in the input sequence. The decomposition guidance weights are set to continuous values ​​from 0 to 1. Preferably, the initial value of the decomposition guidance weights corresponding to the basic load subsequence is set to 0.7, the initial value of the decomposition guidance weights corresponding to the fluctuating load subsequence is set to 0.5, and the initial value of the decomposition guidance weights corresponding to the abrupt load subsequence is set to 0.9. The structure-guided adjustment structure applies the decomposition guidance weights to the decomposition ratio adjustment process of each recursive decomposition layer, ensuring that the basic load subsequence occupies a higher proportion in the decomposition results of the trend change component, that the fluctuating load subsequence occupies a higher proportion in the decomposition results of the fluctuating change component, and that the decomposition ratio is further enhanced at the time position corresponding to the abrupt load subsequence.

[0101] The prediction output structure performs aggregation processing on the multi-level decomposed sequence in time index order, weights and superimposes the trend change components and fluctuation change components in the multi-level decomposed sequence to generate a prediction representation sequence, and outputs a load prediction sequence based on the prediction representation sequence. The load prediction sequence maintains the same time index structure as the input sequence.

[0102] In this embodiment, the improved SCINet model is based on the SCINet model. The main reason for this improvement is that microgrid load time series exhibit significant non-stationarity, multi-timescale superposition, and frequent switching of operating scenarios. Simply relying on the recursive sequence decomposition method of the original SCINet model makes it difficult to distinguish the contribution of different load structures to changes in different operating scenarios during the decomposition process. This can easily lead to coupling shifts between trend change components and fluctuation change components in complex scenarios, thus affecting the stability and consistency of the load prediction sequence. Therefore, this embodiment, while retaining the advantages of the recursive sequence decomposition structure of the SCINet model, introduces a load structure perception guidance mechanism. This mechanism jointly models the load structure identifiers corresponding to the basic load subsequence, fluctuating load subsequence, and abrupt load subsequence with the load operating scenario labels. During the recursive sequence decomposition process, decomposition guidance weights are dynamically generated and embedded into the decomposition ratio adjustment path of each recursive decomposition layer. This enables the proportion of different load structures participating in the decomposition of trend change components and fluctuation change components in different operating scenarios to have adaptive adjustment capabilities. This improvement does not change the original recursive decomposition framework, but introduces dual constraints of structure and scenario at the decomposition ratio level to achieve directional guided decomposition of the inherent structural features of the load time series. It is a structural improvement at the model decomposition mechanism level, with clear technical means, a clear action path and a stable implementation method. Compared with the existing load prediction model that only performs recursive decomposition based on the numerical features of the time series, it has significant improvements in load structure identification accuracy, scenario adaptability and prediction result consistency.

[0103] In this embodiment, S4 specifically refers to:

[0104] The load structure subsequences are input into the improved SCINet model in a unified time index order. The load structure subsequences include the basic load structure subsequence, the fluctuating load structure subsequence, and the abrupt load structure subsequence. Each time index position contains the load power value, the load structure identifier, and the load operation scenario label. A multidimensional input sequence is constructed by time index alignment so that the load power value, the load structure identifier, and the load operation scenario label are arranged in a consistent time dimension.

[0105] Based on the completed input sequence, the basic load subsequence, fluctuating load subsequence, and abrupt load subsequence are distinguished and processed within the improved SCINet model. Through the source relationship of long-term change components, medium- and short-term change components, and abrupt change components formed in the load structure decomposition results, a load structure identifier sequence representing the load structure category is generated. The load structure identifier sequence is represented in discrete numerical form. Preferably, the basic load structure identifier is set to 1, the fluctuating load structure identifier is set to 2, and the abrupt load structure identifier is set to 3. The load structure identifier sequence is arranged in time index order.

[0106] Based on the load structure identifier sequence and load operation scenario label, the decomposition guidance weight is calculated in time index order. The decomposition guidance weight is obtained by mapping the load structure identifier and load operation scenario label to continuous weight values. The weight value range is limited to 0 to 1. Preferably, the decomposition guidance weight corresponding to the basic load is set to 0.6 to 0.8 in the stable operation scenario, the decomposition guidance weight corresponding to the fluctuating load is set to 0.4 to 0.6 in the fluctuating operation scenario, and the decomposition guidance weight corresponding to the sudden load is set to 0.8 to 1.0 in the sudden operation scenario. The decomposition guidance weight sequence is constructed in time index order.

[0107] During the recursive sequence decomposition process, the decomposition guiding weight sequence is introduced into the recursive sequence decomposition calculation path. In each recursive decomposition layer, the decomposition ratio of trend change component and fluctuation change component is adjusted according to the decomposition guiding weight, so that the basic load structure occupies a higher weight in the trend change component and the fluctuation load structure occupies a higher weight in the fluctuation change component. At the same time, the decomposition ratio adjustment amplitude is strengthened at the time position corresponding to the sudden load structure, thereby forming a recursive decomposition output sequence constrained by both the load structure and the operating scenario.

[0108] The recursive decomposition output sequence obtained after the recursive sequence decomposition process is processed by sequence aggregation according to the time index order. The trend change component and fluctuation change component obtained from different recursive decomposition layers are weighted and superimposed. The weight allocation follows the time distribution characteristics of the decomposition guiding weight sequence to generate a load forecast sequence that maintains time continuity.

[0109] In this embodiment, S5 specifically refers to:

[0110] Based on the load forecast sequence and a preset time scale set, the load forecast sequence is resampled according to the time index. The preset time scale set includes minute-level time scales, hour-level time scales, and day-level time scales. The time scale resampling process is completed by executing time aggregation rules on the load forecast sequence within the corresponding time scale range. The time aggregation rules adopt a weighted average method of load forecast values ​​within the time window, and the weights are evenly distributed according to the time index interval. Preferably, the time window length corresponding to the minute-level time scale is set to 5, the time window length corresponding to the hour-level time scale is set to 60, and the time window length corresponding to the day-level time scale is set to 1440, thereby generating a multi-time scale load forecast sequence.

[0111] The generated multi-timescale load forecast sequences are aligned according to a unified time index. By mapping the forecast subsequences at different time scales to a common set of reference time indices, the minute-level, hour-level, and day-level forecast subsequences are aligned in the time dimension, thus constructing a time-scale aligned forecast sequence. The time-scale aligned forecast sequence contains forecast results at different time scales at the same time index position.

[0112] Based on the time-scale aligned prediction sequence, the prediction sub-sequences of different time scales are subjected to consistency verification processing according to the time aggregation rules. The consistency verification processing is completed by comparing the numerical differences between the prediction results of different time scales at the same time index position. The numerical differences are expressed in the form of the absolute difference of the prediction results. Preferably, when the difference between the minute-level prediction result and the hour-level prediction result is greater than 0.15 of the hourly load average, or the difference between the hourly prediction result and the daily prediction result is greater than 0.2 of the daily load average, it is determined that there is a consistency deviation at the corresponding time index position, and thus a consistency deviation sequence is generated along the time index order.

[0113] Based on the consistency deviation sequence, consistency constraint processing is performed on the time-scale aligned prediction sequence. The consistency constraint processing is completed by adjusting the weight distribution of prediction results at different time scales at the locations where consistency deviation exists. At time index locations with large consistency deviations, the weight of prediction results at high-deviation time scales is reduced, while the weight of prediction results at low-deviation time scales is increased. The magnitude of the weight adjustment is positively correlated with the value of the consistency deviation. Preferably, the weight adjustment ratio is set to a range of 0.1 to 0.3. The full sequence constraint calculation is completed through the time index sequence to generate a consistent load prediction sequence.

[0114] In this embodiment, S6 specifically refers to:

[0115] Obtain the actual load sequence that matches the time index of the consistent load prediction sequence. The actual load sequence is derived from the load power data collected in real time during the operation of the microgrid. The actual load sequence is organized according to the same sampling period as the consistent load prediction sequence so that the actual load power values ​​are arranged continuously in the time dimension.

[0116] The consistent load forecast sequence and the actual load sequence are aligned according to the time index. By mapping the consistent load forecast sequence and the actual load sequence to the same time index set, data points with missing or incomplete time indexes are removed, so that the forecast load value and the actual load power value form a pair of data at the same time position, thus forming the forecast load sequence and the actual load sequence.

[0117] The difference calculation process is performed based on the predicted load sequence and the actual load sequence. The difference calculation process is completed by calculating the numerical difference between the predicted load value and the actual load power value at the same time index position. The numerical difference is represented by the absolute value of the difference between the predicted load value and the actual load power value. The load prediction error sequence is generated in the order of time index. Preferably, each error value in the load prediction error sequence is retained to 3 decimal places.

[0118] Based on the load forecasting error sequence, forecasting deviation data is constructed in chronological order. The forecasting deviation data includes a time index, the corresponding forecasted load value, the corresponding actual load power value, and the load forecasting error value. The forecasting deviation data forms a continuous data structure in chronological order.

[0119] In this embodiment, S7 specifically refers to:

[0120] Acquire prediction deviation data and construct a deviation sample sequence based on the prediction deviation data. The deviation sample sequence is organized in time index order. Each time position in the deviation sample sequence contains the corresponding predicted load value, actual load power value and load prediction error value. The deviation sample sequence maintains the continuity consistent with the load time series in the time dimension, so that the change characteristics of the prediction deviation in the time dimension are completely preserved.

[0121] The deviation sample sequence is input into the recursive sequence decomposition structure and prediction output structure inside the improved SCINet model. Inside the model, based on the magnitude and distribution of the load prediction error in the deviation sample sequence, the parameters involved in the recursive sequence decomposition and prediction output calculation are adjusted. The parameter adjustment process is completed by reducing the prediction offset of the time position corresponding to the load prediction error. Preferably, the parameter adjustment range is scaled according to the proportion of the load prediction error value, and the scaling factor is set to 0.05 to 0.2.

[0122] The decomposition guiding weights are updated based on the prediction deviation data. The update of the decomposition guiding weights is accomplished by mapping the load prediction error value to the decomposition guiding weight at the corresponding time index position. This causes the decomposition guiding weight corresponding to the time position with a larger load prediction error to be adjusted in magnitude. Preferably, when the load prediction error value is greater than 0.1, the decomposition guiding weight at the corresponding time position is increased by 10%, and when the load prediction error value is less than 0.05, the decomposition guiding weight at the corresponding time position is decreased by 5%.

[0123] The model parameters after parameter adjustment and the decomposition guiding weights after weight update are written into the improved SCINet model, so that the improved SCINet model updates its internal state while keeping the original structure unchanged. Based on the updated model parameters and decomposition guiding weights, the load prediction sequence is regenerated to form the updated load prediction sequence.

[0124] Example 1:

[0125] To verify the feasibility of this invention in practice, it was applied to a microgrid demonstration station integrating power generation, grid, load, and storage in a riverside industrial park. The demonstration station includes 2.6MW of rooftop photovoltaic capacity, a 1.2MW / 2.4MWh energy storage system, and a mix of production and office loads connected to the system, plus charging loads from electric forklifts and commuter vehicles. The daily load exhibits a typical pattern of "fluctuations in production during the day + sudden changes in charging in the evening + stable base load at night." Existing forecasting methods primarily rely on direct modeling of single load time series. When encountering shift changes, concentrated equipment start-ups and shutdowns, or concentrated charging pile connections, the forecast curve is prone to over-smoothing or peak shifts, leading to increased errors in energy management's judgment of energy storage charging and discharging times, resulting in untimely peak shaving or over-charging and discharging. This embodiment addresses this problem by integrating the method of this invention into the load forecasting chain of the park's daily energy management.

[0126] Load power data from metering terminals is received on the platform side, with a sampling period of 1 minute. Time index breaks are eliminated by aligning sampling gaps with the most recent valid values. Load power values ​​are then mapped to a unified numerical range to form standardized load data, and a load time series is constructed in chronological order. Based on the load time series, it is segmented by time windows (60 seconds in length). The magnitude and rate of change of load power values ​​at adjacent time points are calculated to obtain load change magnitude and rate of change sequences. On this basis, load structure decomposition is performed, dividing the load time series into long-term, medium- and short-term, and abrupt change components. These are then reorganized in chronological order to form a base load subsequence, a fluctuating load subsequence, and an abrupt load subsequence, ensuring that base load, fluctuations, and abrupt changes are separated at the data structure level and can be perceived independently. Simultaneously, load operation scenario identification extracts load power numerical sequences according to time windows, calculates load change amplitude indicators, load change rate indicators, and load level indicators to form a load operation feature sequence, and then performs operation status discrimination processing to obtain a load operation status sequence, generating load operation scenario labels corresponding to the load time series. The scenario label sequence and the load time series use the same time index, and are bound to the labels through time slice mapping, binding scenario label fragments to the basic load subsequence, fluctuating load subsequence, and abrupt load subsequence respectively, and combining them to generate a load structure subsequence. The above data enters the improved SCINet model, where the recursive sequence decomposition structure performs multi-level recursive sequence decomposition on the input sequence, and the structure-guided adjustment structure introduces a load structure perception guidance mechanism, generating decomposition guidance weights based on the load structure identifier and load operation scenario labels, and adjusting the decomposition ratio of each recursive decomposition layer, so that the trend change component accounts for a higher proportion in stable scenarios, the fluctuation change component accounts for a higher proportion in fluctuating scenarios, and the decomposition ratio adjustment amplitude for abrupt time positions is stronger in abrupt scenarios. The predicted output structure aggregates the multi-level decomposition sequences to obtain the load prediction sequence. The load forecast sequence is resampled according to a preset time scale set, which includes three time scales: 5 min, 15 min, and 60 min. The time aggregation rule adopts a weighted average of the forecast values ​​within the window with uniform weight distribution to generate a multi-time scale load forecast sequence. The multi-time scale load forecast sequence is then aligned according to the time index to construct a time scale aligned forecast sequence. A consistency check is performed to obtain a consistency deviation sequence, and consistency constraints are applied to the time scale aligned forecast sequence based on the consistency deviation sequence to form a consistent load forecast sequence.In actual operation, the actual load sequence for the same period is acquired synchronously and aligned with the consistent load prediction sequence. The difference is calculated to generate the load prediction error sequence, which, together with the predicted load sequence and the actual load sequence, constructs the prediction deviation data. The prediction deviation data enters the adaptive update process. The deviation sample sequence is input into the improved SCINet model in time index order to perform parameter update processing. At the same time, the decomposed guiding weights are updated and written into the model, and the updated load prediction sequence is output, enabling the model to continuously track recent load structure changes in the park.

[0127] To reflect the authenticity of the data and the difference in results, this embodiment uses a three-month daily rolling forecast as the verification period, with a forecast step size of 60 minutes in the future. The evaluation indicators are MAE, RMSE, and MAPE. It is compared with the original LSTM method in the park and the SCINet method without the introduction of load structure perception guidance mechanism. The statistical results are shown in Table 1.

[0128] Table 1. Comparison of Three-Month Rolling Forecast Accuracy

[0129]

[0130] Table 1 shows that, based on three months of real-world data, the method of this invention consistently outperforms LSTM and SCINet in terms of MAE, RMSE, and MAPE. Taking September 2025 as an example, the park experienced multiple shift changes coupled with concentrated charging. LSTM achieved a MAPE of 8.36% during the evening charging spike period, demonstrating a lag in tracking the spike load. SCINet showed improvement over LSTM, but still exhibited a problem of small spike amplitude when mixed with fluctuating and spike loads. In contrast, the method of this invention decomposes the load structure into basic load subsequences, fluctuating load subsequences, and spike load subsequences, and associates the load operation scenario labels with the load subsequences to obtain the load structure subsequence. This allows the improved SCINet model to adjust the decomposition ratio based on the decomposition-guided weights during the recursive sequence decomposition process, reducing the MAPE during spike periods to 4.35% and the overall MAPE to 2.91%. This indicates that the modeling of the distinction between spikes and fluctuations improves the sensitivity of short-term predictions while maintaining the stability of overall accuracy.

[0131] In addition to accuracy, the park pays more attention to the consistency of predictions across multiple time scales and the stability of scheduling. In this embodiment, the prediction differences at three time scales of 5 min, 15 min, and 60 min under the same time index are statistically analyzed. The consistency level is represented by the "mean / maximum of the consistency deviation sequence". The "number of charge-discharge switching" and "number of over-limit events (power over-limit or SOC over-limit)" caused by prediction deviations in energy storage scheduling are also statistically analyzed. The comparison results are shown in Table 2.

[0132] Table 2. Comparison of Consistency and Scheduling Stability Across Multiple Time Scales

[0133]

[0134] Table 2 shows that after consistency constraint processing, the divergence between multi-timescale predictions significantly converges using the method of this invention. The mean of the consistency deviation sequence decreases from 52.7kW in SCINet to 31.4kW, and the maximum value decreases from 160.9kW to 96.3kW. This indicates that after consistency verification and consistency constraints, the short-scale predictions do not significantly cancel or amplify the deviations between the long-scale and short-scale predictions. This improvement in consistency is directly reflected on the scheduling side: the number of charge / discharge switching times decreases from 7.8 times / day to 5.1 times / day, the number of over-limit events decreases from 8 times / month to 3 times / month, and the average daily peak shaving contribution of energy storage increases to 2285kWh / day. This indicates that the prediction curve is smoother and closer to the actual load change rhythm, and energy storage operations no longer frequently chase highs and lows. Furthermore, combining the prediction deviation data formed in S6 with the adaptive update process in S7, the model parameters and decomposition guidance weights can be continuously adjusted according to the changes in the park's production rhythm, reducing the mean of the prediction deviation data to 59.8kW, demonstrating the steady-state improvement effect of adaptive updates on long-term operation.

[0135] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A microgrid load forecasting method based on deep learning, characterized in that, Includes the following steps: S1. Collect microgrid load data, perform time alignment and scale normalization processing, and generate load time series; S2. Perform load structure decomposition based on the load time series to generate load subsequences; S3. Based on the load time series, perform load operation scenario identification, generate load operation scenario labels, and associate the load operation scenario labels with the load subsequence to generate a load structure subsequence; S4. Input the load structure subsequence into the improved SCINet model. The improved SCINet model introduces a load structure awareness guidance mechanism and outputs a load prediction sequence. S5. Construct a multi-time-scale load forecast sequence based on the load forecast sequence, and perform consistency constraint processing on the multi-time-scale load forecast sequence to generate a consistent load forecast sequence. S6. Compare the consistent load prediction sequence with the corresponding actual load sequence to generate prediction deviation data; S7. Based on the prediction deviation data, perform adaptive updates on the improved SCINet model and output the updated load prediction sequence.

2. The microgrid load forecasting method based on deep learning according to claim 1, characterized in that, Specifically, S1 is: The load power data generated during the operation of the microgrid is collected, and the load power data is time-aligned according to a unified sampling period to form a load data set with continuous time index; The load data set that has completed time alignment is subjected to scale normalization processing to map the load power values ​​to a uniform value range, generating standardized load data. Based on standardized load data that has undergone scale normalization, a load time series is constructed in chronological order.

3. The microgrid load forecasting method based on deep learning according to claim 1, characterized in that, Specifically, S2 is: The load time series is segmented according to a preset time window, and the load change amplitude sequence and load change rate sequence are calculated based on the load power values ​​at adjacent time points. Based on the load change amplitude sequence and load change rate sequence, load structure decomposition processing is performed on the load time series to divide the load time series into long-term change components, medium- and short-term change components and abrupt change components. The long-term variation components are recombined in chronological order to generate a basic load subsequence; the medium- and short-term variation components are recombined in chronological order to generate a fluctuating load subsequence; and the abrupt change components are recombined in chronological order to generate an abrupt load subsequence. The load subsequences include a basic load subsequence, a fluctuating load subsequence, and an abrupt load subsequence.

4. The microgrid load forecasting method based on deep learning according to claim 1, characterized in that, The process of identifying load operation scenarios based on load time series and generating load operation scenario labels specifically involves: The load time series is segmented according to a preset time window, and the load power numerical sequence within the corresponding time window is extracted. Based on the load power numerical sequence, load change amplitude index, load change rate index and load level index are calculated to construct a load operation characteristic sequence; Based on the load operation feature sequence, the load time series is subjected to operation status discrimination processing to generate a load operation status sequence; Based on the load operation status sequence, load operation scenario labels corresponding to the load time sequence are generated in chronological order.

5. The microgrid load forecasting method based on deep learning according to claim 1, characterized in that, The step of associating the load operation scenario label with the load subsequence to generate the load structure subsequence specifically involves: A scenario tag sequence is constructed for the load operation scenario tags in chronological order, and the scenario tag sequence and the load time series use the same time index. Based on the consistency of time index between the scene label sequence and the load time series, time slice mapping is performed on the scene label sequence to divide the scene label sequence into scene label segments with the same segment boundaries as the load time series. Based on the consistency of the time index between the scene tag fragment and the basic load subsequence, fluctuating load subsequence, and abrupt load subsequence, tag binding processing is performed on the basic load subsequence, fluctuating load subsequence, and abrupt load subsequence respectively to generate the basic load structure subsequence, fluctuating load structure subsequence, and abrupt load structure subsequence. The load structure subsequence is generated by combining the basic load structure subsequence, the load structure subsequence with fluctuations, and the load structure subsequence with abrupt changes.

6. The microgrid load forecasting method based on deep learning according to claim 1, characterized in that, The improved SCINet model includes an input processing structure, a recursive sequence decomposition structure, a structure-guided adjustment structure, and a prediction output structure, specifically: The input processing structure is used to receive the load structure sub-sequence and form an input sequence according to the time index; The recursive sequence decomposition structure is used to perform multi-level recursive sequence decomposition on the input sequence to generate a multi-level decomposed sequence. The structure-guided adjustment structure introduces a load structure perception guidance mechanism. Based on the load structure identifiers corresponding to the basic load subsequence, fluctuating load subsequence and abrupt load subsequence, and combined with the load operation scenario label, a decomposition guidance weight is generated. The decomposition guidance weight is then applied to the recursive sequence decomposition structure to adjust the decomposition ratio of each recursive decomposition layer. The prediction output structure is used to perform aggregation processing on the multi-level decomposition sequence output by the recursive sequence decomposition structure in chronological order to generate a load prediction sequence.

7. The microgrid load forecasting method based on deep learning according to claim 1, characterized in that, Specifically, S4 is: The load structure subsequences are input into the improved SCINet model in time index order to construct the input sequence; Based on the input sequence, the corresponding load structure identification information is extracted from the basic load subsequence, fluctuating load subsequence and abrupt load subsequence in the improved SCINet model to form a load structure identification sequence corresponding to the time index; Based on the load structure identifier sequence and load operation scenario label, the decomposition guidance weights are calculated in chronological order to construct a decomposition guidance weight sequence that is consistent with the time index of the model input sequence. During the recursive sequence decomposition process, the decomposition guiding weight sequence is introduced into the recursive sequence decomposition calculation path, and the decomposition ratio of trend change component and fluctuation change component in each recursive decomposition layer is adjusted to generate the recursive decomposition output sequence. The recursive decomposition output sequence obtained after the recursive sequence decomposition process is processed by sequence aggregation in chronological order to generate the load prediction sequence.

8. The microgrid load forecasting method based on deep learning according to claim 1, characterized in that, Specifically, S5 is: Based on the load forecast sequence and the preset time scale set, the load forecast sequence is resampled according to the time aggregation rule to generate a multi-time scale load forecast sequence. The multi-time scale load forecast sequence includes the forecast sub-sequences corresponding to the preset time scale set. Alignment processing is performed on multi-timescale load forecast sequences according to time index to construct time-scale aligned forecast sequences; Based on the time-scale aligned prediction sequence, consistency verification is performed on the prediction sub-sequences at different time scales according to the time aggregation rules to generate a consistency deviation sequence; Consistency constraint processing is performed on the time-scale aligned prediction sequence based on the consistency deviation sequence to generate a consistent load prediction sequence.

9. A microgrid load forecasting method based on deep learning according to claim 1, characterized in that, Specifically, S6 is: Obtain the actual load sequence that matches the time index of the consistent load forecast sequence; The consistent load forecast sequence and the actual load sequence are aligned according to the time index to form the forecast load sequence and the actual load sequence; A load prediction error sequence is generated by performing difference calculation processing based on the predicted load sequence and the actual load sequence. Prediction bias data are constructed based on the load forecasting error sequence in chronological order.

10. A microgrid load forecasting method based on deep learning according to claim 1, characterized in that, Specifically, S7 is: Obtain forecast deviation data and construct a deviation sample sequence based on the forecast deviation data. The deviation sample sequence includes the forecast load sequence, the actual load sequence, and the load forecast error sequence. The biased sample sequence is input into the improved SCINet model, and parameter update processing is performed based on the biased sample sequence to generate updated model parameters; Based on the prediction deviation data, the decomposed guiding weights are updated to generate the updated decomposed guiding weights. The updated model parameters and the updated decomposition-guided weights are written into the improved SCINet model, and the updated load prediction sequence is output.

Images