Deep learning-based power load multi-period prediction method and system

Abstract

The application discloses a kind of based on deep learning electric power load multi-period prediction method and system, it utilizes instance-level reversible distribution adaptation mechanism, and the original tensor is stationary mapping in input end dynamic extraction local statistics, convert time-varying distribution into unified feature space that model is easy to converge.On this basis, design time-frequency double-flow interactive network, and extract local mutation feature in time domain and global periodicity feature in frequency domain in parallel, and compensate the information missing of single perspective through cross-domain gate fusion mechanism.Finally, cooperate with future-oriented generative decoder and distribution recovery module, and the stationary prediction result is reversely mapped back to the original dimension space with physical meaning.Through this closed-loop logic of stationary reasoning first and distribution restoration later, effectively eliminate the negative influence of statistical drift on model generalization, realize high-precision multi-period prediction under the premise of retaining environment-driven physical characteristics.

Description

Technical Field

[0001] This application relates to the fields of power system automation and data processing technology, and more specifically, to a method and system for multi-period power load prediction based on deep learning. Background Technology

[0002] Multi-period power load forecasting, as a core component of smart grid dispatching, spot market trading, and peak-shaving and valley-filling strategy formulation, is crucial for ensuring the supply-demand balance and operational safety of the power system. With the high proportion of distributed energy resources and the increasing complexity of social electricity consumption patterns, power load data exhibits significant nonlinearity and volatility. This has led to deep learning models, such as Long Short-Term Memory (LSTM) networks and Transformers, gradually replacing traditional statistical methods as the mainstream technology in this field due to their powerful feature extraction capabilities. However, despite the excellent performance of deep learning models in fitting historical data, they still face significant technical challenges when dealing with long-term, multi-period forecasting tasks in real-world industrial scenarios.

[0003] In existing deep learning prediction schemes, models are typically built based on the assumption that training and test data are independent and identically distributed, requiring the input data to remain statistically stationary. However, real-world power load data is inherently non-stationary. With seasonal changes, fluctuations in economic activity, or the impact of sudden weather events, the temporal distribution characteristics of the data (such as mean and variance) dynamically shift, leading to a severe "distribution shift" problem in the inference phase and consequently reducing generalization accuracy. Specifically, existing technologies often employ static normalization methods based on global statistics (such as Z-score standardization), using fixed means and standard deviations to uniformly scale all time steps. This rigid preprocessing logic cannot detect covariate shifts within the load sequence, especially during extreme high temperatures or cold waves, when the load baseline value undergoes drastic physical shifts. In such cases, static normalization often misjudges these environment-driven critical load pulses as statistical noise and smooths them out, resulting in the loss of the physical trend characteristics inherent in the data. This lack of awareness of dynamic inflection points in the distribution not only makes it difficult for the model to capture the real physical driving logic, but also leads to serious lag and distortion in extreme value prediction scenarios, failing to meet the actual needs of the power grid for highly sensitive and robust prediction.

[0004] Therefore, an optimized deep learning-based multi-period power load forecasting scheme is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for multi-period power load forecasting based on deep learning.

[0006] According to one aspect of this application, a deep learning-based method for multi-period power load forecasting is provided, comprising: S1: The acquired raw sensor data stream is imputed for missing values ​​and time-series aligned for sampling to obtain the original spatiotemporal feature tensor. The raw sensor data stream includes historical power load data, meteorological environment data, and time-coded features. S2: Perform invertible instance distribution adaptation on the original spatiotemporal feature tensor to obtain the normalized feature tensor and distribution statistics parameters; S3: Decouple the normalized feature tensor from the time-frequency domain features to obtain the time-domain feature components and the frequency-domain feature components; S4: Perform dual-stream feature interaction and depth extraction on time-domain and frequency-domain feature components to obtain the dual-stream interaction context; S5: Perform multi-stage generative decoding on the dual-stream interactive context to obtain normalized prediction values; S6: Based on the distribution statistics parameters, reverse spatial scaling and physical constraint value smoothing are performed on the normalized forecast values ​​to obtain the power load forecast sequence.

[0007] According to another aspect of this application, a deep learning-based multi-period power load forecasting system is provided, comprising: The multi-source spatiotemporal feature tensor construction module is used to perform missing value interpolation and time-series aligned sampling on the acquired raw sensor data stream to obtain the raw spatiotemporal feature tensor. The raw sensor data stream includes historical power load data, meteorological environment data, and time-coded features. The reversible instance distribution adaptation module is used to perform reversible instance distribution adaptation on the original spatiotemporal feature tensor to obtain the normalized feature tensor and distribution statistics parameters. The time-frequency domain feature decoupling module is used to decouple the normalized feature tensor from the time-frequency domain features to obtain the time-domain feature components and the frequency-domain feature components. The dual-stream feature interaction and depth extraction module is used to perform dual-stream feature interaction and depth extraction on time-domain feature components and frequency-domain feature components to obtain the dual-stream interaction context; The multi-period generative decoding module is used to perform multi-period generative decoding on the dual-stream interaction context to obtain normalized prediction values; The distribution restoration and final output module is used to perform inverse spatial scaling and physical constraint value smoothing truncation on the normalized prediction values ​​based on distribution statistical parameters to obtain the power load prediction sequence.

[0008] Compared with existing technologies, this invention proposes a deep learning-based method for multi-period power load prediction. First, it constructs a spatiotemporal feature tensor containing multi-source information. Utilizing a reversible instance distribution adaptation mechanism, it dynamically removes non-stationary statistical properties from the input data, aligning heterogeneous distributions to a unified normalized feature space. Subsequently, within the model, a time-frequency domain decoupling strategy is employed to extract local abrupt changes in the time domain and global periodic features in the frequency domain in parallel, achieving deep cross-domain interaction through a dual-stream gating mechanism. In the decoding stage, a generative network outputs normalized prediction values ​​for multiple future periods in a single pass. Furthermore, it utilizes previously stored distribution statistics parameters to perform inverse spatial scaling and physical constraint truncation, accurately restoring the prediction results to their original physical dimensions. This effectively eliminates the interference of distribution drift on model generalization while accurately capturing both local fluctuations and long-term trends. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a deep learning-based multi-period power load forecasting method according to an embodiment of this application; Figure 2 This is a data flow diagram of a deep learning-based multi-period power load forecasting method according to an embodiment of this application; Figure 3 This is a flowchart illustrating the process of adapting the original spatiotemporal feature tensor to a normalized feature tensor and distribution statistics parameters by performing reversible instance distribution adaptation on the original spatiotemporal feature tensor according to the deep learning-based multi-period power load prediction method of this application. Figure 4 This is a flowchart illustrating how the distribution stabilization mapping of the original spatiotemporal feature tensor is performed based on the instance mean and instance standard deviation to obtain the normalized feature tensor according to the deep learning-based multi-period power load prediction method of this application. Figure 5 This is a flowchart illustrating the process of decoupling the time-frequency domain features of a normalized feature tensor to obtain time-domain feature components and frequency-domain feature components according to the deep learning-based multi-period power load prediction method of this application. Figure 6 This is a flowchart illustrating the process of performing dual-stream feature interaction and deep extraction on time-domain and frequency-domain feature components to obtain the dual-stream interaction context in the deep learning-based multi-period power load prediction method according to embodiments of this application. Figure 7 This is a block diagram of a deep learning-based multi-period power load prediction system according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules in the systems according to embodiments of this application, any number of different modules can be used and run on power dispatch center server clusters, substation / distribution station edge computing gateways, power consumption management terminals on the consumer side, and / or cloud platforms. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] In current power load forecasting technologies, due to the significant non-stationarity and time-varying nature of load data, traditional normalization methods based on global static statistics are ill-suited to the distribution drift caused by sudden weather changes or seasonal shifts. This dynamic change in distribution characteristics often leads to severe distribution offset challenges for deep learning models during the inference phase, making it difficult for the model to distinguish between real physical driving features and background statistical noise, thereby reducing the generalization accuracy and robustness of multi-period forecasts. Therefore, this application proposes a deep learning-based multi-period power load forecasting method. This method achieves accurate modeling of non-stationary load data by constructing a dual-flow forecasting architecture with adaptive distribution awareness. Specifically, firstly, a reversible instance distribution adaptation is performed on the multi-source original spatiotemporal feature tensor. The non-stationary input is mapped to a unified stationary feature space using dynamically calculated instance mean and standard deviation, effectively isolating distribution drift caused by environmental disturbances. Subsequently, a time-frequency domain feature decoupling mechanism is introduced in the stationary space. Through parallel one-dimensional convolution and fast Fourier transform, the data is decomposed into a time-domain component representing local mutations and a frequency-domain component representing global periods. On this basis, dilated causal convolution and Fourier neural operators are used to deeply extract dual-stream features, and a cross-domain gating mechanism is used to generate a dual-stream interaction context that integrates spatiotemporal multidimensional information. Finally, a generative decoder maps future time periods in one go and performs inverse spatial scaling and physical boundary constraints in combination with previously saved distribution statistics parameters, accurately restoring the normalized prediction values ​​to the original physical dimensions. This ensures data distribution consistency while achieving high-precision and high-reliability prediction of power load under complex operating conditions.

[0016] Figure 1 This is a flowchart of a deep learning-based multi-period power load prediction method according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating a deep learning-based multi-period power load forecasting method according to an embodiment of this application. Figure 1 and Figure 2As shown, the deep learning-based multi-period power load prediction method according to an embodiment of this application includes the following steps: S1, performing missing value interpolation and time-series aligned sampling on the acquired original sensor data stream to obtain an original spatiotemporal feature tensor, the original sensor data stream including historical power load data, meteorological environment data, and time-coded features; S2, performing reversible instance distribution adaptation on the original spatiotemporal feature tensor to obtain a normalized feature tensor and distribution statistics parameters; S3, performing time-frequency domain feature decoupling on the normalized feature tensor to obtain time-domain feature components and frequency-domain feature components; S4, performing dual-stream feature interaction and deep extraction on the time-domain feature components and frequency-domain feature components to obtain a dual-stream interaction context; S5, performing multi-period generative decoding on the dual-stream interaction context to obtain a normalized prediction value; S6, based on the distribution statistics parameters, performing inverse spatial scaling and physical constraint value smoothing truncation on the normalized prediction value to obtain a power load prediction sequence.

[0017] Specifically, in step S1, the acquired raw sensor data stream is imputed for missing values ​​and time-series aligned for sampling to obtain the original spatiotemporal feature tensor. The raw sensor data stream includes historical power load data, meteorological environment data, and time-coded features. It should be noted that due to the inconsistent acquisition frequencies of multi-source heterogeneous sensors in the power IoT environment and the unavoidable packet loss during remote signal transmission, the directly acquired raw data stream often suffers from timestamp misalignment and missing key values. This data discreteness and heterogeneity prevent deep learning models from accurately capturing the continuous dependencies of time-series data, thus interfering with the stability of feature extraction. Therefore, the technical solution of this application first imputes missing values ​​and aligns the acquired raw sensor data stream for sampling to obtain the original spatiotemporal feature tensor. The raw sensor data stream includes historical power load data, meteorological environment data, and time-coded features, thereby constructing a standardized multidimensional input space with unified time granularity and a complete data structure. Through the above processing, spatiotemporal misalignment interference between multi-source data can be effectively eliminated, information gaps caused by transmission failures can be repaired, and a high-fidelity and strictly time-aligned data foundation can be provided for subsequent deep learning models.

[0018] More specifically, in a specific example of this application, step S1 includes: performing feature separation on the original sensor data stream to obtain historical power load data, meteorological environment data, and time-coded features; performing interpolation and alignment processing on the historical power load data and meteorological environment data to obtain an aligned feature group; and performing tensor merging on the aligned feature group and the time-coded features to generate an original spatiotemporal feature tensor.

[0019] More specifically, a structural decoupling operation is performed on the raw data stream of mixed transmission. Feature separation is performed on the raw sensor data stream to obtain historical power load data, meteorological environment data, and time-coded features. The composite data packets are decomposed into feature sequences of independent dimensions for targeted processing. Subsequently, differentiated data cleaning and synchronization strategies are implemented based on the sampling characteristics of different physical quantities. Interpolation and alignment processing is performed on historical power load data and meteorological environment data to obtain aligned feature groups. For discrete missing points in historical power load data caused by equipment failure or network fluctuations, a linear interpolation algorithm is used to smoothly fill in the missing points based on the effective values ​​of adjacent time points. For meteorological environment data, whose sampling frequency is usually lower than that of load data, a combination of resampling and forward padding is used to force alignment to the high-frequency sampling time points of power load data, ensuring strict synchronization between the two in the time dimension. Finally, a multidimensional matrix is ​​constructed for the neural network to read. The aligned feature groups and time-coded features are merged into tensors to generate the original spatiotemporal feature tensor. The completed load sequence, the synchronized temperature and humidity meteorological sequence, and the generated time embedding vector containing daily and weekly periodic information are concatenated and spliced ​​in the channel dimension to form a spatiotemporal feature tensor containing all historical state information and without missing values.

[0020] Specifically, in step S2, the original spatiotemporal feature tensor is reversibly adapted to an instance distribution to obtain a normalized feature tensor and distribution statistics. It should be noted that, given the significant non-stationarity of power load data, and the fact that in real industrial environments, when the power grid encounters extreme temperatures or sudden weather changes, the load baseline value will experience a drastic physical shift. If the traditional globally unified static normalization logic is still used, key load pulse characteristics will be treated as meaningless statistical noise and erased, resulting in severe over-smoothing of the normalization. This lack of awareness of distribution drift inflection points directly makes it difficult for the model to capture the true physical driving logic. Based on this, the technical solution of this application further performs reversible instance distribution adaptation on the original spatiotemporal feature tensor to obtain a normalized feature tensor and distribution statistics. By independently calculating statistics in the time dimension of each sample instance, the non-stationary distribution of the original data is dynamically mapped to a unified stationary feature space, while simultaneously encapsulating distribution metadata for subsequent physical dimension recovery. Through the above processing, a reversible transformation from non-stationary sampling to stationary feature description can be effectively achieved. While preserving the key physical peaks of the load, the negative impact of background distribution shift on model parameter convergence is successfully isolated, thereby enhancing the model's generalization ability and prediction robustness under long-tailed load distribution conditions such as unknown seasons and abnormal weather.

[0021] Figure 3This is a flowchart illustrating the process of reversibly adapting the original spatiotemporal feature tensor to a normalized feature tensor and distribution statistics parameters using a deep learning-based multi-period power load prediction method according to embodiments of this application. Figure 3 As shown, step S2 includes: S21, performing instance statistical analysis on the original spatiotemporal feature tensor in the time step dimension to obtain the instance mean and instance standard deviation; S22, performing distribution stabilization mapping on the original spatiotemporal feature tensor based on the instance mean and instance standard deviation to obtain a normalized feature tensor; S23, encapsulating the instance mean and instance standard deviation with distribution metadata to obtain distribution statistical parameters.

[0022] In step S21, instance statistical analysis is performed on the original spatiotemporal feature tensor along the time step dimension to obtain the instance mean and instance standard deviation. It should be noted that, due to the dual influence of variable meteorological environment and random social electricity consumption behavior, power load data exhibits a non-stationary distribution characteristic that dynamically drifts over time. Furthermore, load values ​​at different sampling periods show objective differences in physical dimensions and fluctuation amplitudes. If feature extraction is directly performed using the original data containing trend terms and different dimensions, the deep model will struggle to adapt to the distribution covariate shift of the input data, leading to oscillations in the gradient descent direction during training and difficulty in capturing the true physical evolution. Therefore, the technical solution of this application further performs instance statistical analysis on the original spatiotemporal feature tensor along the time step dimension to obtain the instance mean and instance standard deviation. This allows for the identification of the central trend and dispersion within the local time window for each independent input sample, constructing benchmark statistical metadata for subsequent non-stationarization mapping. Through the above processing, the relative fluctuation features stripped of specific numerical values ​​can be effectively extracted, providing a consistent input space for the model, and preserving accurate inverse transformation parameters for the restoration of physical dimensions of subsequent prediction results.

[0023] More specifically, in a specific example of this application, a pre-calculation logic of instance normalization is adopted. For each sample instance in the original spatiotemporal feature tensor, statistical operations are performed along the time axis. First, the length of the historical review window covered by the sample is locked. The power load and environmental feature values ​​of all time steps within the window are traversed. The instance mean, which can characterize the basic energy level of the sequence, is calculated by arithmetic averaging. This mean reflects the DC component or background trend under the current scenario. Next, based on the instance mean, the second central moment of each sampling point within the window relative to the mean is further calculated. By accumulating the squares of the differences between the values ​​of each time step and the mean and performing average square root processing, the instance standard deviation, which can quantify the intensity of the fluctuation of the sequence, is obtained. In this calculation process, a very small numerical constant is introduced to avoid the numerical calculation risk of the denominator being zero. Finally, the calculated mean and standard deviation are encapsulated into distribution statistical parameters that correspond one-to-one with the input samples. These parameters not only serve as calculation factors for subsequent distribution stabilization mapping, but also as basic variables for subsequent dynamic compensation and correction combined with scene sensitivity factors.

[0024] In step S22, the original spatiotemporal feature tensor is subjected to distribution stabilization mapping based on the instance mean and instance standard deviation to obtain a normalized feature tensor. It should be noted that in real industrial environments, when the power grid encounters extreme high temperatures or sudden weather changes, the load baseline value will undergo drastic physical shifts. If only static global statistics are used for normalization, key load pulse characteristics will be treated as meaningless statistical noise and erased, resulting in severe over-smoothing of the normalization. This lack of awareness of distribution drift inflection points directly makes it difficult for the model to capture the true physical driving logic. Based on this, the technical solution of this application further performs distribution stabilization mapping on the original spatiotemporal feature tensor based on the instance mean and instance standard deviation to obtain a normalized feature tensor. This process first deconstructs a scene sensitivity factor characterizing the contribution of external factors to distribution drift based on meteorological environmental data and time-coded features. Then, linear compensation is performed on the instance mean, and exponential scaling is performed on the instance standard deviation to obtain the corrected real-time mean and corrected real-time standard deviation. Finally, this dynamically corrected parameter is used to perform adaptive sensitivity stabilization mapping on the original spatiotemporal feature tensor to eliminate the influence of dimensions and fluctuation ratios and remove the local trend term. Through the above processing, an accurate conversion from non-stationary sampling to stationary feature description can be effectively achieved. Mathematically, the negative impact of background distribution shift on model parameter convergence is successfully isolated, while fully preserving the physical mutation characteristics driven by the environment, thereby enhancing the prediction robustness under long-tailed load distribution conditions such as unknown seasons and abnormal weather.

[0025] Figure 4This is a flowchart illustrating the process of performing a distribution stabilization mapping on the original spatiotemporal feature tensor to obtain a normalized feature tensor based on the instance mean and instance standard deviation, according to the deep learning-based multi-period power load prediction method of this application. Figure 4 As shown, step S22 includes: S221, determining the scene sensitivity factor based on meteorological environmental data and time coding features; S222, performing statistical dynamic compensation correction on the instance mean and instance standard deviation based on the scene sensitivity factor to obtain the corrected real-time mean and corrected real-time standard deviation; S223, performing adaptive sensitivity stabilization mapping on the original spatiotemporal feature tensor based on the corrected real-time mean and corrected real-time standard deviation to obtain the normalized feature tensor.

[0026] In step S221, a scene sensitivity factor is determined based on meteorological environmental data and time-coded features. It should be noted that, given the close spatiotemporal heterogeneity between power load data and external environmental factors, if only static mapping logic is used, when the power grid encounters extreme high temperatures or sudden weather changes causing physical shifts in the load baseline value, the globally unified normalization parameters will incorrectly treat key load pulse characteristics as statistical noise and remove them, making it difficult for the model to capture the true physical driving logic. Therefore, the technical solution of this application further determines a scene sensitivity factor based on meteorological environmental data and time-coded features. This allows for the deconstruction of the vulnerability of the current load state to changes in the external environment from multi-dimensional auxiliary information, establishing a quantitative mapping bridge from environmental disturbances to load distribution fluctuations. Through the above processing, the normalization process can be endowed with the ability to perceive external physical laws, making it no longer a simple statistical numerical calculation. This enables accurate differentiation between random fluctuations and physical driving drift, avoiding peak prediction failure due to excessive smoothing.

[0027] More specifically, in a particular example of this application, the execution process involves mapping the meteorological environment data and time-coded features at the current sampling time to a continuous real number space using a fully connected neural network, and then compressing it using a nonlinear activation function to generate a scene sensitivity factor specific to each sampling point, as shown below: in, This represents the scene sensitivity factor generated at time t, with a value ranging from (0,1). It is used to characterize the contribution of external factors to the distribution drift. and These represent the weight matrix and bias parameters of the sensitivity mapping layer, respectively. This represents the environmental characteristics at a given time, such as measured values ​​of temperature and humidity. This represents time-coded information, such as daily cycle coordinates or weekday identifiers. This represents the concatenation operator for feature dimensions. Here, mathematical transformations are used to fuse multidimensional heterogeneous environmental and temporal data into a single scalar signal. This signal serves as a dynamic gating value, intuitively reflecting the degree to which the load forecast is dependent on environmental changes at the current moment. For example, during a sudden heat wave in the afternoon of summer, the temperature values ​​in the meteorological data rise sharply, and the scene sensitivity factor calculated by the neural network approaches 1, indicating that the current load fluctuation is mainly driven by environmental factors, requiring high-intensity dynamic compensation of the distribution parameters. On nights in the mild and stable spring and autumn seasons, this factor may approach 0, indicating that the load fluctuation mainly stems from the randomness of user behavior, and the model maintains relatively stable baseline distribution parameters.

[0028] In step S222, based on the scene sensitivity factor, the instance mean and instance standard deviation are dynamically compensated and corrected using statistics to obtain the corrected real-time mean and corrected real-time standard deviation. It should be noted that, given that global statistics cannot represent the true distribution at each time point when encountering drastic environmental changes, parameters calculated solely using static windows are insufficient to capture the physical drift of power load caused by sudden changes in weather conditions, resulting in the model losing crucial trend information during the normalization stage. Therefore, the technical solution of this application further uses the scene sensitivity factor to dynamically compensate and correct the instance mean and instance standard deviation using statistics to obtain the corrected real-time mean and corrected real-time standard deviation. The scene sensitivity factor generated in the previous step is used to linearly compensate the instance mean, and the instance standard deviation is exponentially scaled and fine-tuned, thereby transforming the static global distribution parameters into real-time distribution parameters that fluctuate with time. Through the above processing, the dynamic psychological expectation adjustment of power plant experts regarding load forecasting benchmarks in the face of sudden weather changes can be simulated, effectively eliminating the interference of non-stationarity on the model input, thus enabling the model to automatically align to the most accurate feature representation space under any meteorological environment.

[0029] More specifically, in a particular example of this application, the execution process involves using external information to dynamically shift the parameters, constructing a mapping relationship from environmental sensitivity to statistical correction through mathematical operations, and obtaining the corrected real-time mean and corrected real-time standard deviation, as shown below: in, This represents the calculated corrected real-time mean, indicating a dynamic benchmark after considering scene disturbances. This represents the global average load level calculated within the original window. This is the scene sensitivity factor output from the preceding sequence. The preset learning compensation coefficient, This is the first-order partial derivative or mean shift vector of the sequence, used to help determine the direction and intensity of compensation. This represents the generated corrected real-time standard deviation. This represents the global fluctuation range under the first embodiment. To control the scaling factor of the sensitivity intervention depth, The function ensures that the compensated standard deviation remains positive and has an exponential adaptive adjustment space. In actual power grid operation scenarios, such as when a sudden thunderstorm in summer causes a sharp drop in temperature, although the global mean calculated based on the past hour remains high, changes in meteorological data will drive the scenario sensitivity factor to increase. Therefore, the linear compensation term in the above formula not only uses the first-order biased guideline to identify the trend direction of the load drop, but also uses the sensitivity factor to dynamically adjust the correction magnitude of the mean, so that the corrected real-time mean can quickly drop to match the current physical conditions. At the same time, the exponential standard deviation correction expands the feature fluctuation tolerance, preventing irregular load fluctuations caused by temperature drops from being misjudged as outliers and truncated by the normalization algorithm, thereby ensuring the physical fidelity of the predicted data under extreme weather conditions.

[0030] In step S223, based on the corrected real-time mean and corrected real-time standard deviation, an adaptive sensitivity stabilization mapping is performed on the original spatiotemporal feature tensor to obtain a normalized feature tensor. It should be noted that, given that the local trend term embedded in the original power load data and the fluctuation ratio that changes over time can severely hinder the feature retrieval efficiency of downstream deep learning networks, and that simple numerical calculations are difficult to retain the physical mutation features driven by the environment while smoothing the data, the technical solution of this application further performs an adaptive sensitivity stabilization mapping on the original spatiotemporal feature tensor based on the corrected real-time mean and corrected real-time standard deviation to obtain a normalized feature tensor, thereby mapping each element of the original spatiotemporal feature tensor to a globally unified named normalized feature space. Through the above processing, a reversible transformation from non-stationary sampling to stationary feature description can be effectively achieved. While retaining the key physical peaks of the load, the negative impact of background distribution shift on model parameter convergence is successfully isolated, enhancing the prediction robustness under long-tailed load distribution conditions such as unknown seasons and abnormal weather.

[0031] More specifically, in a concrete example of this application, an adaptive sensitivity stabilization mapping is performed to transform the original tensor into input features that can be efficiently retrieved by downstream deep learning networks and possess distributional consistency. The process involves subtracting each element of the original spatiotemporal feature tensor from its corresponding corrected real-time mean to remove the local trend term, followed by division to eliminate the influence of units and fluctuation ratios, generating a globally uniformly named normalized feature tensor, represented as follows: in, This indicates the final output after stabilization processing and is the core feature for the model to learn. The raw load data to be processed is input at time t. and These are the dynamically corrected parameters output from step two. This step mathematically constructs an adaptive preprocessing layer to ensure the consistency of the input data in statistical distribution. For example, in a scenario where a cold wave causes a surge in regional electricity load, if traditional static normalization is used, the absolute value of the surge will be compressed into a normal range, thus losing its abnormal warning significance. However, using the method in this embodiment, since the real-time standard deviation after correction at the denominator has been exponentially adapted according to the sensitivity of the scenario, the calculated normalized feature tensor not only eliminates the drift of the baseline value, but also accurately preserves the pulse shape of the sudden event in the relative numerical space, enabling the subsequent model to clearly identify this weather-driven load mutation, rather than misinterpreting it as a normal fluctuation.

[0032] In step S23, the instance mean and instance standard deviation are encapsulated with distribution metadata to obtain distribution statistical parameters. It should be noted that since the feature data after de-stationarization has been stripped of its original physical dimensions and absolute numerical information, retaining only relative fluctuation characteristics, if these key distribution statistics are lost or confused during the layer-by-layer transmission of the deep network, the final prediction result cannot be restored to the megawatt-level power value actually required for power grid dispatch, rendering the prediction model useless for engineering applications. Therefore, the technical solution of this application further encapsulates the instance mean and instance standard deviation with distribution metadata to obtain distribution statistical parameters, thereby constructing a distribution information preservation mechanism independent of the feature extraction stream, providing an accurate mathematical benchmark for subsequent inverse physical recovery. Through the above processing, the full-process traceability of the original data distribution characteristics can be effectively maintained, ensuring that the corresponding statistics can be accurately invoked for inverse spatial mapping during the generative decoding stage, thus achieving lossless regression of the prediction result from the abstract feature space to the real physical space.

[0033] More specifically, in a concrete example of this application, the structured packaging of distribution parameters and tensor dimension alignment are performed. First, the instance mean and instance standard deviation, which are independently calculated in the previous steps, are shape-checked to ensure that they strictly maintain index consistency with the input tensor in the sample batch dimension. Then, the two statistical variables representing the original distribution characteristics are integrated into a unified distribution metadata object by channel cascading or tuple encapsulation. In this process, the metadata object is explicitly marked as a static tensor that does not participate in gradient descent updates and is transmitted in parallel as bypass information to the final output layer of the network. This allows it to bypass the complex intermediate time-frequency feature interaction network to avoid numerical perturbation, ultimately forming distribution statistical parameters that can be directly read and parsed by the subsequent distribution restoration module. This ensures that each prediction time step can be indexed to its corresponding original benchmark and fluctuation amplitude, thereby completing high-precision dimensional restoration.

[0034] Specifically, in step S3, the normalized feature tensor is decoupled from its time-frequency domain features to obtain time-domain and frequency-domain feature components. It should be noted that, given that power load data embeds high-frequency noise generated by random user behavior and low-frequency periodic trends dominated by diurnal rhythms, these multi-scale features are highly coupled in the original single time-series view. If deep feature extraction is directly performed using data mixed with noise and trends, the network often struggles to achieve a balance between capturing local abrupt changes and maintaining long-range global dependencies, easily leading to overfitting of the model to high-frequency noise and obscuring the true long-term evolutionary patterns. Based on this, the technical solution of this application further decouples the normalized feature tensor from its time-frequency domain features to obtain time-domain and frequency-domain feature components. By performing parallel execution of temporal semantic linear projection and discrete spectrum space transformation, the single-dimensional normalized load sequence is physically decomposed into a time-domain stream representing local evolutionary logic and a frequency-domain stream representing global periodic patterns. Through the above processing, a dual-view feature space with positive interactive complementation can be effectively constructed. While using spectral analysis to filter out random noise interference, it provides clear local temporal semantics and global frequency domain fingerprints for subsequent models, thereby improving the model's ability to analyze and reconstruct multi-scale coupling features in complex load curves.

[0035] Figure 5 This is a flowchart illustrating the process of decoupling the time-frequency domain features of a normalized feature tensor to obtain time-domain and frequency-domain feature components, according to the deep learning-based multi-period power load prediction method of this application. Figure 5 As shown, step S3 includes: S31, obtaining time-domain feature components by performing temporal semantic linear projection on the normalized feature tensor; S32, obtaining the original frequency domain spectrum by performing discrete spectral space transformation on the normalized feature tensor; S33, obtaining the frequency domain feature components by performing frequency domain feature extraction and compression processing on the original frequency domain spectrum.

[0036] In step S31, temporal semantic linear projection is performed on the normalized feature tensor to obtain the temporal feature components. It should be noted that although the normalized feature tensor achieves a consistent statistical distribution, its essence remains at a low-dimensional scalar numerical level, lacking high-dimensional semantic information capable of representing the complex temporal logic of power load. Furthermore, directly inputting the original signal into a deep network can easily cause the model to get stuck in local extrema during the shallow feature extraction stage, failing to effectively capture the deep temporal patterns hidden behind numerical fluctuations. Based on this, the technical solution of this application further obtains the temporal feature components by performing temporal semantic linear projection on the normalized feature tensor. A learnable linear mapping matrix is ​​used to project the low-dimensional original observation space to a high-dimensional hidden feature space, thereby constructing a temporal representation foundation with rich semantic information for subsequent dual-stream interaction extraction. Through the above processing, the dimensionality transformation of data from the physical numerical domain to the abstract semantic domain can be effectively realized. While preserving the original sequence time step and causal order, the expressive power of the feature components for local load change trends and dependencies between adjacent time steps is enhanced.

[0037] More specifically, in a specific example of this application, a linear embedding operation in the high-dimensional feature space is performed. For each time step of the normalized feature tensor, a weighted linear combination of power load, meteorological data, and time encoding in the input channel dimension is performed using a weight matrix of a preset dimension or a one-dimensional convolutional kernel. This process is equivalent to a fully connected layer sliding on the time axis, mapping the input vector containing multi-source information into a sequence of latent vectors with unchanged length but expanded depth. Through this end-to-end linear transformation, the originally discrete physical quantities are transformed into continuous embedded vector representations. The generated temporal feature components not only smooth out the minor noise caused by high-frequency acquisition in terms of numerical values, but also initially aggregate the multi-dimensional state information of the current moment at the semantic level, providing a standardized input format for subsequent temporal convolutional networks to capture long-distance dependencies.

[0038] In step S32, the normalized feature tensor undergoes a discrete spectral space transformation to obtain the original frequency domain spectrum. It should be noted that since power load data is essentially a complex signal composed of a superposition of low-frequency periodic components governed by human activity patterns and high-frequency noise components affected by random disturbances, deep neural networks struggle to accurately extract the long-term globally dependent periodic patterns from the aliased time-series waveforms if analyzed only in a single time coordinate system. They are easily affected by local instantaneous fluctuations and lose control over the overall load trend. Therefore, the technical solution of this application further performs a discrete spectral space transformation on the normalized feature tensor to obtain the original frequency domain spectrum. This maps the discrete load sequence in the time domain to the frequency domain through orthogonal transformation, converting the implicit periodic patterns into explicit spectral energy distributions. Through this processing, the signal energy can be effectively refocused in the frequency domain, providing a clear physical perspective for subsequent separation of global periodic features across all time periods and filtering out high-frequency random noise, thereby enhancing the model's analytical accuracy regarding long-term load evolution patterns.

[0039] More specifically, in a specific example of this application, a spectrum mapping operation based on Fast Fourier Transform is performed. For each feature channel in the normalized feature tensor, a one-dimensional Discrete Fourier Transform algorithm is applied along the time step dimension to convert the original real-valued load and environmental feature sequence that changes over time into a complex-valued frequency domain representation composed of real and imaginary parts. In this process, the amplitude and phase information of the input sequence at different frequency points are calculated to generate the original frequency domain spectrum covering the entire frequency band from the DC component to the Nyquist frequency. This spectrum completely preserves all the structured information of the original signal in the frequency domain in the form of a complex tensor, enabling subsequent neural network layers to directly extract and learn the periodicity and phase shift of the load data in the frequency dimension.

[0040] In step S33, frequency domain feature extraction and compression are performed on the original frequency domain spectrum to obtain frequency domain feature components. It should be noted that although the original frequency domain spectrum generated by Fourier transform completely preserves the energy distribution of the signal across the entire frequency band, the main periodic patterns of power load data are usually concentrated in the low-frequency region, while the high-frequency region is often dominated by random noise or transient disturbances. If the full-band information is directly used for subsequent calculations, it will not only introduce a large amount of redundant data and increase computational overhead, but the mixing of high-frequency noise will also interfere with the model's identification and fitting of the core periodic patterns. Based on this, the technical solution of this application further performs frequency domain feature extraction and compression on the original frequency domain spectrum to obtain frequency domain feature components. Through learnable parameterized mapping and selective retention mechanisms, the dominant frequency components with significant energy contributions in the spectrum are extracted, and non-critical high-frequency components carrying noise are filtered out. Through the above processing, the sparsity representation of frequency domain data and the improvement of signal-to-noise ratio can be effectively achieved. While reducing the feature dimension, it ensures that the subsequent model can focus on the key periodic features representing the long-term evolution of the load, thereby improving the operating efficiency and noise robustness of the prediction model.

[0041] More specifically, in a concrete example of this application, feature filtering and dimensionality reduction based on the complex domain are performed. First, a fully connected layer or linear projection layer with complex weights is constructed to map the original frequency domain spectrum containing real and imaginary parts to the hidden layer feature space. Through complex multiplication, the frequency components in the spectrum are weighted and combined and the features are reconstructed to enhance the expressive power of key modes. Subsequently, a Top-K filtering strategy or low-pass filtering truncation logic is adopted to sort the frequency components according to the amplitude intensity after transformation. Only the top K low-frequency main modes with the largest amplitude are retained, or the high-frequency part exceeding the preset threshold is directly truncated. The remaining sparse frequency feature vectors are used as compressed frequency domain feature components. These components, in a compact data form, have undergone denoising and compression, accurately summarizing the global changing trends such as daily and weekly cycles in the load curve, providing high-quality global context input for subsequent deep interaction with time-domain features.

[0042] Specifically, in step S4, dual-stream feature interaction and deep extraction are performed on the time-domain feature components and frequency-domain feature components to obtain the dual-stream interaction context. It should be noted that single-feature-domain extraction methods often present an irreconcilable contradiction between capturing local transient changes and modeling global long-range dependencies. For example, while convolutional networks excel at capturing sudden local load changes caused by weather events, they struggle to cover long-term patterns. Global attention mechanisms, while capable of establishing long-distance correlations, tend to ignore high-frequency local details and have high computational complexity. Therefore, the technical solution of this application further performs dual-stream feature interaction and deep extraction on the time-domain and frequency-domain feature components to obtain the dual-stream interaction context. This allows for the use of time-domain convolutional dependency enhancement processing to deeply mine the local short-term lag effects of sequences, the use of spectral refinement and spatial domain restoration processing to capture global periodic trends in the frequency dimension, and the adaptive complementary fusion of the two feature streams through a cross-domain gating fusion interaction mechanism. Through the above processing, the locality limitation of spatiotemporal feature extraction can be effectively broken, and a high-dimensional dual-flow context integrating microscopic physical mutations and macroscopic periodic laws can be constructed, thereby improving the model's joint analysis accuracy and expression completeness of long and short time series features in complex and ever-changing power grid operation environment.

[0043] Figure 6 This is a flowchart illustrating the process of performing dual-stream feature interaction and deep extraction on time-domain and frequency-domain feature components to obtain the dual-stream interaction context in the deep learning-based multi-period power load prediction method according to embodiments of this application. Figure 6 As shown, step S4 includes: S41, performing temporal convolution dependency enhancement processing on the temporal feature components to obtain enhanced temporal features; S42, performing spectral refinement and spatial restoration processing on the frequency domain feature components to obtain enhanced frequency domain features; S43, performing cross-domain gating fusion interaction on the enhanced temporal features and enhanced frequency domain features to obtain a dual-stream interaction context.

[0044] In step S41, temporal convolutional dependency enhancement processing is performed on the temporal feature components to obtain enhanced temporal features. It should be noted that since the power load sequence contains a large number of instantaneous pulse features caused by sudden weather conditions or the start-up and shutdown of large equipment, these nonlinear local abrupt changes often carry crucial short-term dependency information. If conventional standard convolutional kernels are used for feature extraction, limited by their fixed geometric structure and limited receptive field, the model struggles to capture large-scale temporal dependencies without significantly increasing network depth, and easily overlooks causal constraints in the time dimension, leading to future information leakage. Based on this, the technical solution of this application further performs temporal convolutional dependency enhancement processing on the temporal feature components to obtain enhanced temporal features, thereby introducing a causal convolutional mechanism with a dilated structure, exponentially expanding the perceptual range of feature extraction while maintaining the temporal resolution. Through the above processing, the model's ability to capture local load abrupt changes and their hysteresis response patterns can be effectively strengthened while strictly adhering to temporal causal logic, providing accurate short-term dynamic context for prediction.

[0045] More specifically, in a concrete example of this application, sequence enhancement operations based on dilated causal convolution are performed. For the temporal feature components of the input, a multi-layered convolutional neural network is constructed. Each layer is configured with a convolutional kernel having a specific dilation factor, which determines the sampling interval stride of the kernel on the input sequence. Zero values ​​are inserted between kernel elements to sparsify the sampling points, thereby expanding the coverage along the time axis while keeping the number of parameters constant. During this process, sliding window convolution calculations are performed, forcibly constraining the kernel to only perform weighted summation of feature points at the current and historical time points, strictly prohibiting the use of information from future time points, thereby generating enhanced temporal features. The mathematical expression of this calculation process is as follows: in, This represents the temporal feature vector enhanced by dilated convolution, which aggregates local temporal information within the expanded receptive field. This represents the learnable weight parameters of the convolution kernel at position k, used to extract the feature importance at different time lag points. This represents the original time-domain feature components of the input. Indicates the current time step index. This represents the dilation factor, used to control the jump span of convolutional kernel sampling. As the number of network layers increases, this factor typically grows exponentially to cover longer historical dependencies. This represents the index variable within the convolution kernel size. This represents the total size of the convolution kernel. Through this mechanism, the model can gradually expand from microscopic dependencies between adjacent time steps to macroscopic dependencies over long periods, accurately identifying key features such as short-term load drops caused by thunderstorms.

[0046] In step S42, the frequency domain feature components undergo spectral refinement and spatial restoration to obtain enhanced frequency domain features. It should be noted that although the initially extracted frequency domain feature components retain the main energy distribution of the load, they still lack adaptive optimization for global periodic patterns, and the spectral features in the pure frequency domain cannot directly interact semantically with local features in the time domain. Therefore, the technical solution of this application further refines and restores the frequency domain feature components to obtain enhanced frequency domain features, thereby performing parameterized filtering and feature enhancement on the spectrum in the complex frequency domain and mapping the global, time-dependent features back to the time-domain spatial coordinate system. Through the above processing, the mathematical property that global convolution in the time domain is equivalent to dot product multiplication in the frequency domain can be effectively utilized to capture long-distance coupling relationships with low computational cost and achieve dimensional alignment with time-domain features for subsequent cross-domain fusion.

[0047] More specifically, in a concrete example of this application, the spectral processing logic based on Fourier neural operators is executed. First, the transmitted frequency domain feature components are received. Then, element-wise multiplication is performed on the spectrum using a learnable complex weight matrix. This process acts as a dynamic filter, adaptively amplifying the frequency components crucial for prediction and suppressing irrelevant background noise. Subsequently, a one-dimensional inverse fast Fourier transform is performed on the processed spectrum to restore the global features from the frequency domain back to the time series space, generating enhanced frequency domain features. The mathematical expression of this calculation process is as follows: in, This represents the enhanced frequency domain characteristics when converted back to the time domain. This represents the one-dimensional inverse fast Fourier transform operator. This represents the spectral filtering weight matrix in the complex domain, used to learn the global dependency pattern of the load. This represents the frequency domain feature components of the main step input. Through this transformation, the originally abstract global periodic trend is transformed into a time-series representation aligned with the time axis.

[0048] In step S43, enhanced time-domain features and enhanced frequency-domain features are subjected to cross-domain gated fusion interaction to obtain a dual-stream interaction context. It should be noted that, since enhanced time-domain features focus on capturing local load pulses and short-term hysteresis effects caused by sudden events, while enhanced frequency-domain features focus on characterizing the global periodic evolution trend throughout the entire time period, the contribution of these two types of features to the final load curve differs objectively at different prediction times. If only simple linear superposition or channel splicing is used for fusion, it is difficult to reflect the dynamic importance of features at different time steps, easily leading to the dilution of key dominant features by secondary features. Based on this, the technical solution of this application further performs cross-domain gated fusion interaction of enhanced time-domain features and enhanced frequency-domain features to obtain a dual-stream interaction context, thereby constructing a context-aware adaptive weight allocation mechanism that dynamically adjusts the information transmission ratio of the time-domain stream and the frequency-domain stream according to the current load state. Through the above processing, the complementary integration of microscopic physical mutation information and macroscopic periodic laws can be effectively achieved, ensuring that the model can automatically align to the optimal feature combination space when facing time-domain dominant scenarios caused by sudden weather or frequency-domain dominant scenarios during stable operation.

[0049] More specifically, in a concrete example of this application, a feature fusion operation based on a nonlinear gating mechanism is performed. First, enhanced time-domain features are used as driving signals, mapped through a linear layer containing weight matrices and bias terms, and then compressed to the range of zero to one by a sigmoid activation function. This generates fusion gating weights that can quantitatively evaluate the relative importance of the time-domain features at the current moment. The logic for generating these weights aims to determine whether the current load state is in a rapidly fluctuating, unsteady state. Subsequently, these gating weights are used as balancing coefficients to perform a complementary element-wise weighted summation of the enhanced frequency-domain features and the enhanced time-domain features, generating a two-stream interactive context. The mathematical expression of this calculation process is as follows: in, This represents the calculated fusion weight, used to dynamically adjust the contributions of the time and frequency channels. This represents the Sigmoid activation function, which maps values ​​to the interval between 0 and 1. and Let represent the weight matrix and bias term of the gated network, respectively. This represents the fused feature context of the final output. Represents element-wise multiplication operators. and These refer to the enhanced time-domain features and enhanced frequency-domain features generated by the preceding steps, respectively. Through this mechanism, when the model detects a significant abrupt change in the enhanced time-domain features, the gating network automatically adjusts the weights to prioritize the time-domain information; otherwise, it focuses on the periodic guidance in the frequency domain.

[0050] Specifically, in step S5, multi-time-period generative decoding is performed on the dual-stream interaction context to obtain normalized predicted values. It should be noted that traditional time series forecasting tasks often employ recursive decoding strategies, using the predicted value from the previous time step as the input for the next. This serial generation method leads to an exponential accumulation of early prediction errors over long-term forecasting scenarios, severely weakening the accuracy of long-term predictions. Furthermore, while the high-dimensional dual-stream interaction context contains rich time-frequency features, directly using it for regression prediction results in significant data redundancy and dimensionality mismatch. Therefore, the technical solution of this application further performs multi-time-period generative decoding on the dual-stream interaction context to obtain normalized predicted values, thereby constructing a parallel generative output mechanism that maps deep abstract features to all future time steps to be predicted at once. Through the above processing, the error propagation risk caused by recursive prediction can be effectively avoided. Simultaneously, feature compression and nonlinear activation improve the model's ability to fit future load change trends, ensuring the consistency of prediction accuracy across all time periods.

[0051] More specifically, in a specific example of this application, step S5 includes: performing feature space compression projection on the two-stream interaction context to obtain a predicted feature vector; performing a future-time-oriented generative fully connected weight mapping on the predicted feature vector to obtain a prediction tensor to be activated; and performing dimensional reconstruction and Gaussian error linear activation on the prediction tensor to be activated to obtain a normalized prediction value.

[0052] More specifically, the deep neural network-based decoding and reconstruction operation first involves feature dimensionality reduction and purification of the high-dimensional dual-stream interaction context. The dual-stream interaction context is then compressed and projected into its feature space to obtain a predicted feature vector. A fully connected layer or linear projection layer is used to map the context tensor, which incorporates time-frequency information, to a low-dimensional latent space, removing collinearity redundancy and retaining the core prediction drivers. Subsequently, a direct mapping from the latent feature space to the future physical time axis is established. Generative fully connected weight mapping oriented towards future time periods is performed on the predicted feature vector to obtain the prediction tensor to be activated. This process does not rely on autoregressive iterative generation but instead uses a dense output layer with shared parameters to calculate the load prediction response for all time points within the specified future time period in one go, generating an intermediate tensor containing information about the future time series length. Finally, the output data is morphologically adapted and nonlinearly smoothed. The tensor to be activated is redimensionally reconstructed and Gaussian error linear activation is applied to obtain normalized prediction values. The data dimensions are adjusted to a sequence format that meets the requirements of downstream tasks through tensor reshaping operations. The predicted values ​​are then randomized and smoothed using the Gaussian error linear unit activation function to simulate the continuous probability distribution characteristics of power load at the micro level, thereby outputting the final prediction sequence within the normalized space.

[0053] Specifically, in step S6, based on distribution statistics parameters, the normalized predicted values ​​are subjected to inverse spatial scaling and physical constraint value smoothing truncation to obtain the power load prediction sequence. It should be noted that since the prediction results output by the generative decoding of the deep neural network are essentially still within a compressed dimensionless normalized feature space, their numerical distribution follows a standard normal distribution rather than the physical power dimensions actually required for power grid dispatch. Furthermore, the deep model may generate negative values ​​or abnormal extreme values ​​that violate physical common sense during pure mathematical fitting. Without targeted physical domain restoration and constraints, the prediction data cannot be directly applied to the supply and demand balance calculation and safety verification of the power system. Based on this, the technical solution of this application further performs inverse spatial scaling and physical constraint value smoothing truncation on the normalized predicted values ​​based on distribution statistics parameters to obtain the power load prediction sequence, thereby establishing an inverse mapping mechanism from the abstract feature space back to the real physical space, and introducing physical boundary conditions to correct the legality of the model output. Through the above processing, the true fluctuation range and reference energy level of the power load data can be effectively restored, and non-physical noise caused by algorithm fitting deviation can be eliminated, thereby ensuring that the final generated load prediction sequence meets the actual requirements of power grid operation in terms of both numerical accuracy and physical logic.

[0054] More specifically, in a specific example of this application, step S6 includes: performing distribution feature recovery on the distribution statistics parameters to obtain the instance mean and instance standard deviation; performing inverse spatial scaling adaptation on the normalized predicted values ​​based on the instance standard deviation and instance mean to obtain the original dimensional predicted values; and obtaining the power load prediction sequence by performing non-negative truncation and physical boundary constraints on the original dimensional predicted values.

[0055] Furthermore, the precise reconstruction and boundary correction of physical dimensions are performed. First, the distribution statistics parameters encapsulated and passed to the output in the preprocessing stage are parsed and unpacked. The distribution characteristics of the distribution statistics parameters are restored to obtain the instance mean and instance standard deviation. The structured metadata tensor is restored to a statistical vector that is strictly aligned with the prediction sequence in the time and batch dimensions. Subsequently, using the inverse operation logic of linear transformation, based on the instance standard deviation and instance mean, the normalized prediction values ​​are subjected to inverse spatial scaling adaptation processing to obtain the original dimensional prediction values. Element-wise multiplication is performed with the instance standard deviation to recover the fluctuation amplitude, and the calculation result is added with the instance mean to restore the load benchmark, thereby mapping the relative value back to the megawatt-level power space. Finally, the physical operation constraint logic of the power system is introduced. By performing non-negative truncation and physical boundary constraints on the original dimensional prediction values, the power load prediction sequence is obtained. Nonlinear activation functions or conditional judgment logic are applied to forcibly filter out load prediction values ​​less than zero, and abnormal peaks that exceed the physical carrying capacity limit of the power grid are smoothly truncated. Finally, a high-confidence prediction result that conforms to physical laws is output.

[0056] Specifically, during the training phase of the deep learning model, a training sample set containing input feature sequences and future target labels is first constructed based on historical power load data. The training sample set is then divided into batches using a sliding window mechanism. Subsequently, the training batch data is input into an initialized dual-stream interactive network to perform forward propagation computation. Based on the aforementioned reversible instance distribution adaptation logic, statistical parameters and normalized features of the training samples are dynamically generated. The normalized prediction values ​​of the training phase are output through time-frequency decoupling and gated fusion paths. Next, a loss function to measure the prediction accuracy is constructed, and the numerical difference between the normalized prediction value and the corresponding real normalized load label is calculated. The gradient information of each learnable parameter in the scene sensitivity mapping layer, time-frequency feature extraction layer, and cross-domain gated fusion layer is calculated based on the chain rule. Finally, the backpropagation algorithm and adaptive optimizer are used to perform end-to-end iterative updates to the network weight matrix and bias terms until the loss function value converges to a preset stable interval. This completes the parameterized learning of the spatiotemporal evolution law of non-stationary power load by the model.

[0057] In summary, the deep learning-based multi-period power load prediction method according to the embodiments of this application is explained. It performs reversible instance distribution adaptation on the original spatiotemporal feature tensor, using dynamically calculated instance statistics to map non-stationary inputs to a unified stationary feature space, effectively isolating distribution drift caused by environmental disturbances. Based on this, a time-frequency domain feature decoupling and a dual-stream deep interaction mechanism are introduced to extract local abrupt changes and global periodic trends in parallel in the time and frequency domains, respectively. A dual-stream interaction context is generated through gated fusion, thus completely preserving the physical evolution law of the load within the stationary space. Finally, a generative decoder is used to map future time periods in one step, and inverse spatial scaling and physical boundary constraints are performed in conjunction with distribution statistics parameters to accurately restore the prediction results to the original dimensions, achieving high-precision and robust multi-period prediction of non-stationary power loads under complex operating conditions.

[0058] Furthermore, a deep learning-based multi-period power load forecasting system is also provided.

[0059] Figure 7 This is a block diagram of a deep learning-based multi-period power load prediction system according to an embodiment of this application. Figure 7 As shown, the deep learning-based multi-period power load prediction system 100 according to an embodiment of this application includes: a multi-source spatiotemporal feature tensor construction module 110, used to perform missing value interpolation and time-series aligned sampling on the acquired original sensor data stream to obtain the original spatiotemporal feature tensor, the original sensor data stream including historical power load data, meteorological environment data, and time-coded features; a reversible instance distribution adaptation module 120, used to perform reversible instance distribution adaptation on the original spatiotemporal feature tensor to obtain the normalized feature tensor and distribution statistics parameters; and a time-frequency domain feature decoupling module 130, used to perform reversible instance distribution adaptation on the original spatiotemporal feature tensor to obtain the normalized feature tensor and distribution statistics parameters; and a time-frequency domain feature decoupling module 130, used to perform time-frequency domain feature decoupling on the normalized feature tensor. The time-frequency domain features are decoupled from the normalized feature tensor to obtain time-domain and frequency-domain feature components; the dual-stream feature interaction and deep extraction module 140 is used to perform dual-stream feature interaction and deep extraction on the time-domain and frequency-domain feature components to obtain the dual-stream interaction context; the multi-time period generative decoding module 150 is used to perform multi-time period generative decoding on the dual-stream interaction context to obtain the normalized prediction value; the distribution restoration and final output module 160 is used to perform inverse spatial scaling and physical constraint value smoothing truncation on the normalized prediction value based on the distribution statistical parameters to obtain the power load prediction sequence.

[0060] As described above, the deep learning-based multi-period power load prediction system 100 according to the embodiments of this application can be implemented in various types of computing devices or control units. For example, it can be deployed in a server cluster of a provincial power dispatch center or installed in an edge computing node of a substation smart gateway. In one possible implementation, the deep learning-based multi-period power load prediction system 100 according to the embodiments of this application can be integrated into the computing device as a software module and / or a hardware module. For example, the system 100 can be a resident data processing service in the operating system of the computing device, and the software module is configured to perform missing value imputation and time-series alignment of multi-source heterogeneous data, reversible distribution adaptation and normalization mapping based on instance statistics, time-frequency dual-domain decoupling and deep interactive extraction for non-stationary features, and generative decoding and physical dimension restoration for future time periods, or it can be a dedicated intelligent power load prediction algorithm program developed for the computing device. Of course, the system 100 can also be one of many hardware modules of the computing device or control unit, or it can be embedded in a field-programmable gate array circuit to accelerate fast Fourier transform and high-dimensional tensor convolution operations in parallel, or it can be a deep neural network inference acceleration integrated circuit for a specific application.

[0061] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for multi-period forecasting of power load based on deep learning, characterized in that, include: S1: The acquired raw sensor data stream is imputed for missing values ​​and time-series aligned for sampling to obtain the original spatiotemporal feature tensor. The raw sensor data stream includes historical power load data, meteorological environment data, and time-coded features. S2: Perform invertible instance distribution adaptation on the original spatiotemporal feature tensor to obtain the normalized feature tensor and distribution statistics parameters; S3: Decouple the normalized feature tensor from the time-frequency domain features to obtain the time-domain feature components and the frequency-domain feature components; S4: Perform dual-stream feature interaction and depth extraction on time-domain and frequency-domain feature components to obtain the dual-stream interaction context; S5: Perform multi-stage generative decoding on the dual-stream interactive context to obtain normalized prediction values; S6: Based on the distribution statistics parameters, reverse spatial scaling and physical constraint value smoothing are performed on the normalized forecast values ​​to obtain the power load forecast sequence.

2. The deep learning-based multi-period power load forecasting method according to claim 1, characterized in that, Step S1 includes: Feature separation is performed on the raw sensor data stream to obtain historical power load data, meteorological environment data, and time-coded features; Historical power load data and meteorological environment data are interpolated and aligned to obtain aligned feature groups; Tensor merging is performed on the aligned feature groups and the time-coded features to generate the original spatiotemporal feature tensor.

3. The deep learning-based multi-period power load forecasting method according to claim 1, characterized in that, Step S2 includes: In the time step dimension of the original spatiotemporal feature tensor, instance statistical analysis is performed on the original spatiotemporal feature tensor to obtain the instance mean and instance standard deviation; Based on the instance mean and instance standard deviation, the original spatiotemporal feature tensor is subjected to a distribution stabilization mapping to obtain a normalized feature tensor. The instance mean and instance standard deviation are encapsulated with distribution metadata to obtain distribution statistics parameters.

4. The deep learning-based multi-period power load forecasting method according to claim 1, characterized in that, Step S3 includes: Temporal feature components are obtained by performing temporal semantic linear projection on the normalized feature tensor; The normalized feature tensor is subjected to discrete spectral space transformation to obtain the original frequency domain spectrum; Frequency domain features are extracted and compressed from the original frequency domain spectrum to obtain frequency domain feature components.

5. The deep learning-based multi-period power load forecasting method according to claim 1, characterized in that, Step S4 includes: Temporal convolution dependency enhancement processing is performed on the temporal feature components to obtain enhanced temporal features; Spectral refinement and spatial restoration of frequency domain feature components are performed to obtain enhanced frequency domain features; Cross-domain gating fusion interaction is performed on enhanced time-domain features and enhanced frequency-domain features to obtain a dual-stream interaction context.

6. The deep learning-based multi-period power load forecasting method according to claim 1, characterized in that, Step S5 includes: The feature space is compressed and projected onto the two-stream interaction context to obtain the predicted feature vector; Perform a future-time-oriented generative fully connected weight mapping on the predicted feature vector to obtain the prediction tensor to be activated; The dimensionality of the prediction tensor to be activated is reconstructed and Gaussian error linear activation is applied to obtain normalized prediction values.

7. The deep learning-based multi-period power load forecasting method according to claim 1, characterized in that, Step S6 includes: The distribution characteristics of the distribution statistics parameters are restored to obtain the instance mean and instance standard deviation; Based on the instance standard deviation and instance mean, the normalized predicted values ​​are subjected to inverse spatial scaling adaptation to obtain the original dimensional predicted values. The power load forecast sequence is obtained by performing non-negative truncation and physical boundary constraints on the original dimensional forecast values.

8. The deep learning-based multi-period power load forecasting method according to claim 3, characterized in that, Based on the instance mean and instance standard deviation, a distribution stabilization mapping is performed on the original spatiotemporal feature tensor to obtain a normalized feature tensor, including: Determine scene sensitivity factors based on meteorological environmental data and time coding characteristics; Based on the scene sensitivity factor, the instance mean and instance standard deviation are dynamically compensated and corrected by statistics to obtain the corrected real-time mean and corrected real-time standard deviation. Based on the corrected real-time mean and corrected real-time standard deviation, an adaptive sensitivity stabilization mapping is performed on the original spatiotemporal feature tensor to obtain a normalized feature tensor.

9. A deep learning-based multi-period power load forecasting system, characterized in that, include: The multi-source spatiotemporal feature tensor construction module is used to perform missing value interpolation and time-series aligned sampling on the acquired raw sensor data stream to obtain the raw spatiotemporal feature tensor. The raw sensor data stream includes historical power load data, meteorological environment data, and time-coded features. The reversible instance distribution adaptation module is used to perform reversible instance distribution adaptation on the original spatiotemporal feature tensor to obtain the normalized feature tensor and distribution statistics parameters. The time-frequency domain feature decoupling module is used to decouple the normalized feature tensor from the time-frequency domain features to obtain the time-domain feature components and the frequency-domain feature components. The dual-stream feature interaction and depth extraction module is used to perform dual-stream feature interaction and depth extraction on time-domain feature components and frequency-domain feature components to obtain the dual-stream interaction context; The multi-period generative decoding module is used to perform multi-period generative decoding on the dual-stream interaction context to obtain normalized prediction values; The distribution restoration and final output module is used to perform inverse spatial scaling and physical constraint value smoothing truncation on the normalized prediction values ​​based on distribution statistical parameters to obtain the power load prediction sequence.

10. The deep learning-based multi-period power load forecasting system according to claim 9, characterized in that, The reversible instance distribution adaptation module includes: The instance statistical analysis unit is used to perform instance statistical analysis on the original spatiotemporal feature tensor in the time step dimension to obtain the instance mean and instance standard deviation. The distribution stabilization mapping unit is used to perform distribution stabilization mapping on the original spatiotemporal feature tensor based on the instance mean and instance standard deviation to obtain a normalized feature tensor. The distribution metadata encapsulation unit is used to encapsulate the instance mean and instance standard deviation into distribution metadata to obtain distribution statistical parameters.

Images