A cross-period load collaborative forecasting method based on deep meteorological learning

Abstract

This invention relates to the field of power system automation technology, specifically to a cross-time period load collaborative forecasting method based on deep meteorological learning, comprising the following steps: S100: acquiring multidimensional meteorological data and historical load data of the target power grid area and surrounding meteorological stations; constructing a meteorological spatiotemporal topology map, and mapping the multidimensional meteorological data to the same time step as the historical load data through a cross-frequency feature alignment mechanism; S200: inputting the meteorological spatiotemporal topology map into a pre-constructed spatiotemporal feature extraction network to extract the dynamic evolution features of the meteorological system in terms of spatial distribution and time lag, and outputting a comprehensive meteorological spatiotemporal feature vector. This invention pioneers a cross-scale dual-branch network and a cross-attention gating mechanism, which perfectly simulates the physical modulation effect of "macro-climate trends" on "micro-local meteorological abrupt changes" at the feature decoupling level, significantly improving the model's prediction robustness under complex extreme weather superposition states.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and specifically to a cross-period load collaborative forecasting method based on deep meteorological learning. Background Technology

[0002] With the rapid development of new power systems, the integration of a high proportion of renewable energy has made the net load of the power grid extremely sensitive to weather conditions. Accurate cross-period power load forecasting (simultaneously outputting short-term micro-dispatch instructions and long-term macro-reserve instructions) has become a core foundation for ensuring the safe and stable operation of the power grid. Currently, load forecasting methods based on deep learning (such as LSTM, Transformer, GCN, etc.) are widely used. However, with the deepening of engineering applications, existing technologies have gradually exposed the following three deep-seated technical shortcomings: 1. The fusion of multi-source heterogeneous meteorological data is awkward and ignores the physical advection and spatiotemporal lag effects of the meteorological system.

[0003] Existing forecasting models typically input data on temperature, humidity, and wind speed from multiple meteorological stations directly into the network after simple concatenation. On one hand, the sampling frequency of meteorological data (e.g., hourly) and load data (e.g., 15-minute) are often inconsistent, and current techniques mostly use linear or spline interpolation for frequency alignment. This purely mathematical smoothing process directly obliterates the nonlinear abrupt changes in sudden weather events (e.g., downpours, cloud cover) at the micro-timescale. On the other hand, existing graph networks often construct static connections based solely on geographical distance, failing to effectively quantify the time lag effects generated when macro-weather systems such as cold air masses and rain belts move across regions. This results in the model's inability to capture the advection and diffusion evolution of meteorological elements in the spatial dimension.

[0004] 2. Temporal feature extraction suffers from "scale entanglement" and lacks physical constraints on the modulation of micro-meteorology by macro-climate.

[0005] The impact of meteorology on load exhibits significant multi-scale characteristics: macro-climate (such as cold waves and droughts) determines the low-frequency trend of base load, while micro-local weather (such as sudden gusts) triggers high-frequency fluctuations in load. However, existing sequence forecasting models typically process input data with a uniform temporal resolution, leading to severe entanglement (i.e., "scale entanglement") between long-term low-frequency trends and short-term high-frequency fluctuations in the feature space. More critically, existing models fail to understand the "physical modulation effect of macro-background on micro-abrupt changes" (e.g., a sudden drop of 2°C in temperature can cause a surge in heating load in a severe winter, but have minimal impact in early autumn), resulting in extremely poor predictive robustness under complex meteorological superposition states.

[0006] 3. The fragmented forecasting across time periods and the lack of physical consistency constraints on the conservation of electrical energy lead to scheduling command conflicts.

[0007] In cross-time-scale collaborative forecasting tasks, existing deep learning frameworks typically treat long-term macro-level forecasts and short-term micro-level forecasts as two independent mathematical regression problems, employing purely data-driven mean squared error (MSE) for independent training. This paradigm completely disregards the physical reality of power systems, ignoring the cumulative conservation characteristic of load as energy over the time domain. In practical engineering, this easily leads to the serious problem of "a huge deviation between the time-domain integral (cumulative value) of the short-term forecast sequence and the total value of the long-term macro-level forecast." This physical inconsistency across time scales directly causes severe command conflicts (i.e., dispatch command fragmentation) when the power grid dispatch center formulates real-time unit start-up and shutdown plans (based on short-term forecasts) and inter-regional power transmission and fuel storage plans (based on long-term forecasts), posing significant safety hazards to power grid operation.

[0008] In summary, there is an urgent need for a cross-period load collaborative forecasting method that can deeply couple the spatiotemporal evolution of meteorology with the consistency constraints of power physics, in order to solve the technical problems of data alignment distortion, characteristic scale entanglement, and conflict between long-term and short-term forecasting instructions in existing technologies. Summary of the Invention

[0009] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a cross-time period load collaborative forecasting method based on deep meteorological learning, comprising the following steps: S100: acquiring multidimensional meteorological data and historical load data of the target power grid area and surrounding meteorological stations; constructing a meteorological spatiotemporal topology map, and mapping the multidimensional meteorological data to a time step consistent with the historical load data through a cross-frequency feature alignment mechanism; S200: inputting the meteorological spatiotemporal topology map into a pre-constructed spatiotemporal feature extraction network to extract the dynamic evolution features of the meteorological system in terms of spatial distribution and time lag, and outputting a comprehensive meteorological spatiotemporal feature vector; S300: fusing the comprehensive meteorological spatiotemporal feature vector with the historical load data and inputting it into... In the cross-scale dual-branch network, long-term low-frequency trend features influenced by macro-climate and short-term high-frequency fluctuation features influenced by sudden weather events are extracted respectively. Based on the cross-scale gating mechanism, the extraction process of short-term high-frequency fluctuation features is dynamically adjusted using the long-term low-frequency trend features. S400: Based on the output of the cross-scale dual-branch network, long-term load forecast values ​​and short-term load forecast sequences are generated respectively. During the network training phase, the network parameters are updated based on the joint loss function containing physical consistency constraints to constrain the time-domain cumulative value of the short-term load forecast sequence to approximate the long-term load forecast value. The trained network outputs cross-time-period collaborative load forecast results.

[0010] Further, in step S100, the construction of the meteorological spatiotemporal topology map specifically includes: defining the target power grid area and surrounding meteorological stations as graph nodes; for any two meteorological station nodes, calculating the inverse of the spatial geographic distance between them, as well as the dynamic time-normalized distance of the historical meteorological element sequence; The inverse of the spatial geographic distance and the dynamic time-normalized distance are weighted and fused to generate a comprehensive association weight between stations; a weight threshold is set, and comprehensive association weights greater than the weight threshold are retained as edge weights, thereby constructing the meteorological spatiotemporal topology map that includes the spatial connectivity and time lag of the meteorological system.

[0011] Furthermore, the cross-frequency feature alignment mechanism specifically includes: constructing a frequency alignment attention module to address the time step difference between the low-frequency sampled multidimensional meteorological data and the high-frequency sampled historical load data; encoding the time step position of the high-frequency sampled data as a query vector, and using the low-frequency sampled multidimensional meteorological data sequence as a key vector and a value vector; By calculating the inner product of the query vector and the key vector and performing normalization, the nonlinear interpolation weights of each high-frequency time step on the surrounding low-frequency meteorological data nodes are obtained; the value vector is weighted and summed using the nonlinear interpolation weights to output a high-frequency meteorological feature sequence that is completely aligned with the historical load data time steps.

[0012] Further, in step S200, the spatiotemporal feature extraction network includes multiple cascaded spatiotemporal convolutional modules; each of the multiple spatiotemporal convolutional modules sequentially includes a gated temporal convolutional layer, a spatial graph diffusion convolutional layer, and a second gated temporal convolutional layer. The extraction of dynamic evolution features of the meteorological system in terms of spatial distribution and temporal lag specifically includes: performing a one-dimensional causal convolution operation on the input meteorological feature sequence through the gated temporal convolution layer to extract the initial temporal evolution features of the meteorological station's own elements; The initial temporal evolution features are input into the spatial graph diffusion convolutional layer. Combined with the edge weights of the meteorological spatiotemporal topology graph, the features of adjacent meteorological stations are spatially aggregated in multiple hops to simulate the spatial advection diffusion process of meteorological elements between different geographical stations and output spatial aggregation features. The spatial aggregation features are input into the second gated temporal convolutional layer, and the temporal receptive field is expanded by dilated causal convolution to extract the time lag effect of meteorological features after spatial diffusion over a longer period.

[0013] Furthermore, the specific process of multi-hop spatial aggregation in the spatial graph diffusion convolutional layer is as follows: a state transition matrix is ​​constructed based on the comprehensive correlation weights of the meteorological spatiotemporal topology graph; The spatial propagation of meteorological elements is modeled as a two-way random walk diffusion process. By using a preset number of diffusion steps, Chebyshev polynomial approximations of different meteorological station nodes under forward and reverse diffusion are calculated, and the graph signals under each number of diffusion steps are weighted and summed to capture global topological features such as the movement of cold air masses or rain belts in the regional power grid.

[0014] Further, in step S300, the cross-scale dual-branch network includes a long-time series extraction branch and a short-time series extraction branch. The extraction of long-term low-frequency trend features influenced by macro-climate and short-term high-frequency fluctuation features influenced by sudden weather events specifically includes: The integrated meteorological spatiotemporal feature vector and the historical load data are concatenated according to time steps to form a fused feature sequence; The fused feature sequence is input into the long-term sequence extraction branch, and a large-step pooling layer is used to perform time-domain downsampling on the fused feature sequence to filter out local high-frequency meteorological noise. The result is then input into the first recurrent neural network to extract the long-term low-frequency trend features with a period of day or week. The fused feature sequence is synchronously input into the short-time sequence extraction branch to maintain the original high-frequency temporal resolution. The short-time high-frequency fluctuation features caused by sudden weather events such as short-term rain or cloud cover are captured by a one-dimensional dilated convolutional network.

[0015] Furthermore, the process of dynamically adjusting the extraction of short-term high-frequency fluctuation features based on the cross-scale gating mechanism specifically includes: constructing a cross-scale attention gating unit to perform time-dimensional upsampling interpolation on the long-term low-frequency trend features so that its sequence length is aligned with the short-term high-frequency fluctuation features; The aligned long-term low-frequency trend features are linearly mapped to generate a query vector. The short-time high-frequency fluctuation features are linearly mapped to generate key vectors and value vectors, respectively. Calculate the scaled dot product of the query vector and the key vector, and activate it using the Softmax function to generate a cross-scale attention weight matrix.

[0016] Furthermore, the cross-scale attention weight matrix characterizes the physical modulation intensity of macro-climate evolution background on meteorological abrupt changes at each micro-time step; multiplying the cross-scale attention weight matrix with the value vector yields the modified short-term features guided by macro-trends, and the residuals of the modified short-term features are added to the short-term high-frequency fluctuation features to output the final collaborative short-term features.

[0017] Furthermore, the physical consistency penalty weight The dynamic weights, which are dynamically and adaptively adjusted with the number of training iterations, have the following adjustment strategy: In the early stages of network training, the following were assigned: and A higher initial value, and the Set to a small value close to 0 to drive the cross-scale dual-branch network to preferentially converge the local prediction accuracy in both long and short time periods; As the number of training iterations increases, the value of the short-time error weight is gradually decayed using the exponential annealing or cosine annealing algorithm. In the later stages of network training, when the decreasing gradient of the mean square error of the short-term load prediction sequence and the mean square error of the long-term load prediction value is lower than a preset convergence threshold, the... By assigning a dominant weight, the integral of the short-term load forecast sequence is forced to strictly approximate the long-term load forecast value, thereby achieving a physical collaborative closed loop across time scales.

[0018] Beneficial effects This invention accurately recreates the dynamic evolution of meteorological elements in spatial advection diffusion and temporal long-tail lag by constructing a meteorological spatiotemporal map and frequency alignment mechanism. Secondly, it pioneers a cross-scale dual-branch network and cross-attention gating mechanism, perfectly simulating the physical modulation effect of "macro-climate trends" on "micro-local meteorological abrupt changes" at the feature decoupling level, significantly improving the model's predictive robustness under complex extreme weather superposition states. More importantly, this invention creatively introduces a joint loss function based on the law of conservation of electrical energy into the backpropagation logic at the model's underlying level. Through a dynamic annealing training strategy, it forces the time-domain cumulative value of the high-frequency short-term prediction sequence to strictly approximate the low-frequency long-term prediction total value, fundamentally eliminating the engineering-level disaster risk of "tears and conflicts between long and short-term instructions" in cross-time grid scheduling. Finally, it outputs a collaborative prediction result with both extremely high mathematical accuracy and strict physical consistency, providing an unbreakable decision-making foundation for the safe, stable, and economical operation of new power systems. Attached Figure Description

[0019] Figure 1 This is a flowchart of a cross-period load collaborative forecasting method based on deep meteorological learning according to the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] The present invention will now be described in further detail with reference to the accompanying drawings: Example: As shown in the figure, a cross-period load collaborative forecasting method based on deep meteorological learning includes the following steps: S100: Acquire multidimensional meteorological data and historical load data of the target power grid area and surrounding meteorological stations; construct a meteorological spatiotemporal topology map, and map the multidimensional meteorological data to a time step consistent with the historical load data through a cross-frequency feature alignment mechanism; S200: Input the meteorological spatiotemporal topology map into the pre-constructed spatiotemporal feature extraction network to extract the dynamic evolution features of the meteorological system in terms of spatial distribution and time lag, and output a comprehensive meteorological spatiotemporal feature vector; S300: The integrated meteorological spatiotemporal feature vector is fused with the historical load data and input into a cross-scale dual-branch network to extract long-term low-frequency trend features affected by macro-climate and short-term high-frequency fluctuation features affected by sudden weather events, respectively; and based on the cross-scale gating mechanism, the extraction process of short-term high-frequency fluctuation features is dynamically adjusted using the long-term low-frequency trend features. S400: Based on the output of the cross-scale dual-branch network, long-term load forecast values ​​and short-term load forecast sequences are generated respectively; during the network training phase, the network parameters are updated based on the joint loss function containing physical consistency constraints, so as to constrain the time-domain cumulative value of the short-term load forecast sequence to approximate the long-term load forecast value, and the trained network outputs cross-time-period collaborative load forecast results.

[0023] Furthermore, the specific implementation process of step S100 is as follows: The system first acquires multidimensional meteorological data, such as temperature, humidity, wind speed, and solar irradiance, from multiple meteorological stations within the target power grid area and its surrounding buffer zone via a data acquisition module. Simultaneously, it acquires historical load data for the power grid area. In this process, those skilled in the art typically face a common engineering challenge: the sampling frequency of meteorological stations is usually low (e.g., once per hour), while the load sampling frequency for micro-level power dispatch is high (e.g., once every 15 minutes). Conventional linear interpolation or spline interpolation methods only smooth the data mathematically, completely ignoring the nonlinear abrupt changes in sudden weather events (such as downpours or cloud cover) at the micro-timescale, thus leading to distortion of underlying features.

[0024] To overcome the aforementioned shortcomings and construct a physically meaningful underlying input, this embodiment creatively introduces a meteorological spatiotemporal topology map construction mechanism. The system abstracts the acquired target power grid area and surrounding meteorological stations as nodes in a graph structure. When determining the edge weights between nodes, this embodiment does not use the conventional single Pearson correlation coefficient, but instead comprehensively considers both static geographic spatial attributes and dynamic meteorological evolution lag attributes. The system first calculates the reciprocal of the spatial geographic distance between any two meteorological station nodes to characterize their proximity in static geographic space. More importantly, to capture time lag effects such as the passage of cold fronts or the shift of rain belts, the system calculates the dynamic time-warped distance (DTW) for the historical meteorological element sequences of these two stations. Compared to the conventional Euclidean distance, the DTW distance can effectively quantify the similarity between two time series with temporal phase shifts, thereby accurately characterizing the physical phenomenon of "the time lag required for meteorological conditions to propagate from station A to station B." Subsequently, the system performs a weighted fusion of the inverse of the spatial geographic distance and the dynamic time-normalized distance to generate a comprehensive association weight between nodes. After setting a specific weight threshold, weakly correlated connections are filtered out, and edges greater than the threshold are retained. This constructs a meteorological spatiotemporal topology map that not only reflects the geographic topology but also contains the dynamic transfer inertia of meteorology.

[0025] After constructing the meteorological spatiotemporal topology map, this embodiment further designs a cross-frequency feature alignment module based on a frequency alignment attention mechanism to address the aforementioned high- and low-frequency data alignment problem. This module abandons traditional numerical interpolation and instead transforms the high-frequency alignment process into an "adaptive intelligent query process." Specifically, the system encodes the time step positions of high-frequency samples (e.g., 15-minute intervals) using a linear mapping as the query vector for the attention mechanism, while mapping the multidimensional meteorological data sequences of low-frequency samples (e.g., hourly intervals) into key and value vectors. During computation, the system calculates the scaled inner product between the high-frequency query vector and the low-frequency key vector, and then normalizes and activates it using the Softmax function, thereby adaptively learning the nonlinear interpolation weights of each high-frequency micro-time step on its surrounding low-frequency meteorological data nodes. This weight matrix essentially reflects the inherent law of nonlinear evolution of meteorological elements at the micro-time scale. Finally, the system uses these nonlinear interpolation weights to perform a weighted summation of the value vectors containing actual meteorological information, outputting a high-frequency comprehensive meteorological feature sequence that is perfectly aligned with the historical load data time step.

[0026] This embodiment not only deeply restores the spatiotemporal physical evolution law of the meteorological system in the data preprocessing stage, but also uses the attention mechanism to achieve lossless high-frequency alignment of heterogeneous data at the micro-time scale. This provides consistent and spatiotemporally dynamic semantic-rich low-level feature inputs for the subsequent cross-scale dual-branch network, which greatly lays the foundation for the high accuracy and robustness of the overall prediction model.

[0027] Furthermore, the specific implementation process of step S200 is as follows: This embodiment aims to address the technical limitation of conventional sequence models (such as LSTM or a single CNN) in simultaneously considering the advection diffusion of meteorological elements in the spatial dimension and the evolutionary lag in the temporal dimension. In the real physical world, abrupt weather changes in a power grid area often do not arise out of thin air, but are caused by the advection of meteorological systems (such as cold air masses or rain belts) from adjacent areas along wind direction and pressure gradients. To accurately characterize this complex dynamic evolution process, this embodiment inputs the aforementioned meteorological spatiotemporal topology map into a pre-constructed spatiotemporal feature extraction network. This network does not use a conventional single network layer, but is composed of multiple cascaded customized spatiotemporal convolutional modules. In particular, to deeply simulate the meteorological evolution patterns, each module adopts a "time-space-time (TST)" sandwich hierarchical architecture, which sequentially includes a gated temporal convolutional layer, a spatial graph diffusion convolutional layer, and a second gated temporal convolutional layer.

[0028] In practice, the first step in feature extraction focuses on the initial evolution of local weather. The system first performs a one-dimensional causal convolution operation on the meteorological feature sequences of each station in the meteorological spatiotemporal topology map through a first-layer gated temporal convolution. The physical significance of this step lies in extracting the initial changing trends of meteorological elements at each independent station over time (e.g., the continuous decrease in air pressure and increase in humidity that a location inevitably experiences before a rainstorm). This causal convolution strictly adheres to the chronological order of events, ensuring that the model does not use future meteorological data to deduce the current state, thus avoiding data leakage problems in prediction tasks.

[0029] Next, to capture the spatial transfer characteristics of the meteorological system, the system inputs local initial temporal evolution features into the core spatial graph diffusion convolutional layer. In this layer, this embodiment abandons the conventional graph convolution operation that can only handle single-hop neighbor nodes, and instead creatively introduces a multi-hop spatial aggregation mechanism based on a bidirectional random walk diffusion process. The system constructs a state transition matrix based on the comprehensive correlation weights of the meteorological spatiotemporal topology graph and uses a preset multi-order Chebyshev polynomial to approximate the frequency domain filtering process of the graph signal. From the perspective of physical meteorology, this multi-order graph diffusion operation perfectly simulates the spatial advection diffusion process of meteorological elements between different geographical stations. Different diffusion steps (i.e., the order of the Chebyshev polynomial) directly map the influence radius and movement trajectory of the meteorological system (such as a strong convective cloud) within the regional power grid space. Through this weighted summation of forward and reverse graph signals, the network can accurately capture the moving inertia of the macro-weather system in the global topology.

[0030] Finally, considering that the impact of meteorological systems on power grid load after spatial transfer often exhibits a long-tailed hysteresis effect (for example, after cold air arrives, the impact of low temperatures on heating load can last for hours or even days), the system further inputs the spatially aggregated features into the second gated temporal convolutional layer. In this layer, this embodiment specifically employs dilated causal convolution. By introducing a dilation factor, the network exponentially expands the model's temporal receptive field without increasing computational parameters or losing temporal resolution. This allows the network to retrospectively analyze and integrate meteorological evolution information over longer historical periods, thereby deeply extracting the time lag effect caused by spatially diffused meteorological features within the target power grid area. After iterative extraction by the above multi-module approach, the system finally outputs a comprehensive meteorological spatiotemporal feature vector containing complete spatial transfer and temporal lag dynamic inertia, providing solid, high-fidelity high-order physical feature support for subsequent cross-time-period joint forecasting.

[0031] Furthermore, the specific implementation process of step S300 is as follows: After obtaining the comprehensive meteorological spatiotemporal feature vector, step S300 of this embodiment aims to solve the core engineering problem of conflicting long-term and short-term instructions in power system dispatching. Traditional deep learning sequence models (such as a single Transformer or LSTM) typically use a uniform temporal resolution to process the input sequence, which physically violates the multi-scale characteristics of meteorological evolution: macro-climate (such as long-period events like cold waves and droughts) are slowly changing low-frequency signals that determine the basic load trend of the power grid; while micro-meteorology (such as localized downpours or sudden cloud cover blocking photovoltaic panels) are rapidly changing high-frequency signals that trigger instantaneous load spikes in the power grid. To achieve physical decoupling between the two during the feature extraction stage...

[0032] First, the system seamlessly concatenates the integrated meteorological spatiotemporal feature vector with historical load data into a fused feature sequence at time steps. This sequence is then synchronously input into parallel long-term and short-term series extraction branches. In the long-term series extraction branch, to effectively filter out "high-frequency meteorological noise" such as regular minute temperature fluctuations from day to night, the system first employs a large-step pooling layer to perform temporal downsampling on the fused feature sequence. The downsampled feature sequence is then input into the first recurrent neural network, using days or weeks as time windows to accurately extract long-term low-frequency trend features deeply influenced by macro-climate and seasonal periodic patterns. These features physically characterize the baseline load evolution profile of the power grid over a future period. Simultaneously, in the short-term series extraction branch, the system strictly preserves the original high-frequency temporal resolution (e.g., 15-minute level) and utilizes the wide receptive field of a one-dimensional dilated convolutional network (1DDilatedCNN) to keenly capture short-term high-frequency fluctuation features caused by sudden severe weather. Through this heterogeneous resolution dual-branch design, the model achieves a perfect separation between macro-climate background and micro-meteorological events in the underlying feature space.

[0033] This can be further understood as follows: the decoupled features do not exist in isolation. In the real physical world, macroscopic climate backgrounds often strongly modulate the destructive power of microscopic meteorological events (for example, a sudden drop in temperature of the same magnitude of 2°C has little impact on the load if it occurs in the macroscopic background of "midsummer," but will cause a sharp surge in heating load if it occurs in the macroscopic background of "severe winter cold wave"). In order to mathematically model this physical modulation phenomenon at the network level, this embodiment further constructs a cross-scale attention gating mechanism between the two branches. In specific implementation, the system first performs time-dimensional upsampling interpolation on the long-term low-frequency trend features to accurately align them with the short-term high-frequency fluctuation features in terms of sequence length.

[0034] Subsequently, the cross-scale gating mechanism generates a query vector (or simply query vector) by linearly mapping the aligned long-term low-frequency trend features. And map the short-term high-frequency fluctuation features into key vectors (Key, abbreviated as Key) respectively. ) and value vector, abbreviated as The system calculates... and The scaled dot product is then activated by the Softmax function to generate a cross-scale attention weight matrix. In this process, the long-term trend is used as... Go check short-term fluctuations Essentially, it allows the AI ​​model to adaptively assess: "Under the current macro-climate context, how much actual impact does a meteorological abrupt change at a specific micro-time step have on the power grid load?" This weight matrix physically and intuitively represents the dynamic modulation intensity of the macro-evolutionary background on abrupt changes at each micro-time step. Finally, the system combines this cross-scale attention weight matrix with the actual value vector containing local fluctuation information. Matrix multiplication is performed to obtain modified short-term features strongly guided by macro trends. These features are then added to the original short-term high-frequency fluctuation features via residual connections, ultimately outputting collaborative short-term features with a global perspective and rich details.

[0035] Furthermore, the specific implementation process of S400 is as follows: In actual power grid operation and dispatching, short-term forecasts (such as load forecasts every 15 minutes over the next 24 hours) are typically used to guide real-time generator start-up and shutdown and frequency regulation; while long-term forecasts (such as the total load forecast for the next day) are used to guide large-scale inter-regional power transmission and fuel storage. Existing deep learning forecasting frameworks typically treat long-term and short-term forecasting tasks as independent mathematical regression problems, using conventional mean squared error (MSE) or cross-entropy as loss functions for independent training. This purely data-driven training paradigm completely ignores the conservation characteristics of power grid load as a physical energy source, which can easily lead to significant deviations between the integral (or cumulative value) of short-term forecast results in the time domain and the long-term forecast results. For example, the sum of 96 15-minute short-term forecasts is far greater than the total load forecast for a single day, which will directly lead to decision conflicts and resource misallocation in the dispatch center.

[0036] To completely eliminate this engineering vulnerability from the underlying algorithmic logic, this embodiment creatively introduces a joint loss function based on energy conservation and multi-scale physical consistency constraints in the joint prediction stage at the network's end. Specifically, the cross-scale dual-branch network generates long-term load forecasts at the output layer. and contain Short-term load forecast sequence at high-frequency time steps During the backpropagation phase of network training, the system does not rely on a single data error to update the learnable parameters of the network model. Instead, a joint loss function was constructed that includes a physical consistency penalty term.

[0037] In this joint loss function, the mean square error of the short-term load forecast sequence and the mean square error of the long-term load forecast are used to measure the model's absolute numerical approximation ability on their respective time scales. This is achieved by combining the physical cumulative value with the macroscopic long-term forecast value. The squared difference is introduced as a penalty term into the loss function. When the model updates the weights using the gradient descent method, it will be forced to find an optimal solution space that not only satisfies the historical data distribution but also strictly adheres to the electrical physics conservation law that "the sum of micro loads equals the total macro load".

[0038] Furthermore, considering that deep learning models are prone to getting stuck in local optima or experiencing gradient explosion when dealing with multi-objective optimization, this embodiment does not use fixed penalty weights. Instead, it employs a dynamic weight strategy that dynamically adjusts weights based on the number of training iterations. In the initial stage of network training, the system assigns weights based on short-term error. and long-term error weights Higher initial values, and physical consistency penalty weights Set to a small value close to 0. The physical significance of this is that it allows the two-branch network to fully learn and fit the weather-load mapping relationship at their respective scales under relatively relaxed physical constraints, ensuring the model possesses basic local predictive capabilities. As the number of training iterations increases, the system gradually decays using a cosine annealing algorithm. and The value of is taken, and it increases exponentially in sync. The value of is determined in the later stages of network training convergence, when the local data error no longer decreases significantly. It occupies the dominant weight in the loss function. This annealing training strategy, which shifts from "relaxed individual learning" to "forced physical convergence," can smoothly and stably drive the integral of the short-term load forecast sequence to strictly approximate the expected long-term forecast value. The final output is a cross-time collaborative forecast result that achieves a high degree of consistency in both mathematical accuracy and physical logic, providing an extremely reliable data foundation for the safe and economical joint dispatch of the power grid.

[0039] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-period load collaborative forecasting method based on deep meteorological learning, characterized in that, Includes the following steps: S100: Acquire multidimensional meteorological data and historical load data of the target power grid area and surrounding meteorological stations; construct a meteorological spatiotemporal topology map, and map the multidimensional meteorological data to a time step consistent with the historical load data through a cross-frequency feature alignment mechanism; S200: Input the meteorological spatiotemporal topology map into the pre-constructed spatiotemporal feature extraction network to extract the dynamic evolution features of the meteorological system in terms of spatial distribution and time lag, and output a comprehensive meteorological spatiotemporal feature vector; S300: The integrated meteorological spatiotemporal feature vector is fused with the historical load data and input into a cross-scale dual-branch network to extract long-term low-frequency trend features affected by macro-climate and short-term high-frequency fluctuation features affected by sudden weather events, respectively. Furthermore, based on a cross-scale gating mechanism, the extraction process of short-term high-frequency fluctuation features is dynamically adjusted using the long-term low-frequency trend features. S400: Based on the output of the cross-scale dual-branch network, generate long-term load forecast values ​​and short-term load forecast sequences respectively; During the network training phase, the network parameters are updated based on a joint loss function that includes physical consistency constraints, so as to constrain the time-domain cumulative value of the short-term load forecast sequence to approximate the long-term load forecast value, and the trained network outputs cross-time-period collaborative load forecast results.

2. The cross-period load collaborative forecasting method based on deep meteorological learning according to claim 1, characterized in that, In step S100, the construction of the meteorological spatiotemporal topology map specifically includes: defining the target power grid area and surrounding meteorological stations as graph nodes; for any two meteorological station nodes, calculating the inverse of the spatial geographic distance between them and the dynamic time-normalized distance of the historical meteorological element sequence; The inverse of the spatial geographic distance and the dynamic time-normalized distance are weighted and fused to generate a comprehensive association weight between stations; a weight threshold is set, and comprehensive association weights greater than the weight threshold are retained as edge weights, thereby constructing the meteorological spatiotemporal topology map that includes the spatial connectivity and time lag of the meteorological system.

3. The cross-period load collaborative forecasting method based on deep meteorological learning according to claim 2, characterized in that, The cross-frequency feature alignment mechanism specifically includes: constructing a frequency alignment attention module to address the time step difference between the low-frequency sampled multidimensional meteorological data and the high-frequency sampled historical load data; encoding the time step position of the high-frequency sampled data as a query vector, and using the low-frequency sampled multidimensional meteorological data sequence as a key vector and a value vector; By calculating the inner product of the query vector and the key vector and performing normalization, the nonlinear interpolation weights of each high-frequency time step on the surrounding low-frequency meteorological data nodes are obtained; the value vector is weighted and summed using the nonlinear interpolation weights to output a high-frequency meteorological feature sequence that is completely aligned with the historical load data time steps.

4. The cross-period load collaborative forecasting method based on deep meteorological learning according to claim 3, characterized in that, In step S200, the spatiotemporal feature extraction network includes multiple cascaded spatiotemporal convolutional modules; each of the multiple spatiotemporal convolutional modules sequentially includes a gated temporal convolutional layer, a spatial graph diffusion convolutional layer, and a second gated temporal convolutional layer. The extraction of dynamic evolution features of the meteorological system in terms of spatial distribution and temporal lag specifically includes: performing a one-dimensional causal convolution operation on the input meteorological feature sequence through the gated temporal convolution layer to extract the initial temporal evolution features of the meteorological station's own elements; The initial temporal evolution features are input into the spatial graph diffusion convolutional layer. Combined with the edge weights of the meteorological spatiotemporal topology graph, the features of adjacent meteorological stations are spatially aggregated in multiple hops to simulate the spatial advection diffusion process of meteorological elements between different geographical stations and output spatial aggregation features. The spatial aggregation features are input into the second gated temporal convolutional layer, and the temporal receptive field is expanded by dilated causal convolution to extract the time lag effect of meteorological features after spatial diffusion over a longer period.

5. The cross-period load collaborative forecasting method based on deep meteorological learning according to claim 4, characterized in that, The specific process of multi-hop spatial aggregation in the spatial graph diffusion convolutional layer is as follows: a state transition matrix is ​​constructed based on the comprehensive correlation weights of the meteorological spatiotemporal topology graph; The spatial propagation of meteorological elements is modeled as a two-way random walk diffusion process. By using a preset number of diffusion steps, Chebyshev polynomial approximations of different meteorological station nodes under forward and reverse diffusion are calculated, and the graph signals under each number of diffusion steps are weighted and summed to capture global topological features such as the movement of cold air masses or rain belts in the regional power grid.

6. The cross-period load collaborative forecasting method based on deep meteorological learning according to claim 5, characterized in that, In step S300, the cross-scale dual-branch network includes a long-time series extraction branch and a short-time series extraction branch. The extraction of long-term low-frequency trend features influenced by macro-climate and short-term high-frequency fluctuation features influenced by sudden weather events, respectively, specifically includes: The integrated meteorological spatiotemporal feature vector and the historical load data are concatenated according to time steps to form a fused feature sequence; The fused feature sequence is input into the long-term sequence extraction branch, and a large-step pooling layer is used to perform time-domain downsampling on the fused feature sequence to filter out local high-frequency meteorological noise. The result is then input into the first recurrent neural network to extract the long-term low-frequency trend features with a period of day or week. The fused feature sequence is synchronously input into the short-time sequence extraction branch to maintain the original high-frequency temporal resolution. The short-time high-frequency fluctuation features caused by sudden weather events such as short-term rain or cloud cover are captured by a one-dimensional dilated convolutional network.

7. The cross-period load collaborative forecasting method based on deep meteorological learning according to claim 6, characterized in that, The process of dynamically adjusting the extraction of short-term high-frequency fluctuation features based on the cross-scale gating mechanism specifically includes: constructing a cross-scale attention gating unit to perform time-dimensional upsampling interpolation on the long-term low-frequency trend features so that its sequence length is aligned with the short-term high-frequency fluctuation features; The aligned long-term low-frequency trend features are linearly mapped to generate a query vector. The short-time high-frequency fluctuation features are linearly mapped to generate key vectors and value vectors, respectively. Calculate the scaled dot product of the query vector and the key vector, and activate it using the Softmax function to generate a cross-scale attention weight matrix.

8. The cross-period load collaborative forecasting method based on deep meteorological learning according to claim 7, characterized in that, The cross-scale attention weight matrix characterizes the physical modulation intensity of macro-climate evolution background on meteorological abrupt changes at each micro-time step; multiplying the cross-scale attention weight matrix with the value vector yields the modified short-term features guided by macro-trends, and the residuals of the modified short-term features are added to the short-term high-frequency fluctuation features to output the final collaborative short-term features.

9. A cross-period load collaborative forecasting method based on deep meteorological learning according to claim 8, characterized in that, The physical consistency penalty weight The dynamic weights, which are dynamically and adaptively adjusted with the number of training iterations, have the following adjustment strategy: In the early stages of network training, the short-term error weight and the long-term error weight are given relatively high initial values, and the physical consistency penalty weight is set to a small value close to 0, so as to drive the cross-scale dual-branch network to converge the local prediction accuracy of long-term and short-term errors first. As the number of training iterations increases, the value of the short-term error weight is gradually reduced using exponential annealing or cosine annealing algorithms. In the later stages of network training, when the descent gradient between the mean square error of the short-term load prediction sequence and the mean square error of the long-term load prediction value is lower than a preset convergence threshold, the physical consistency penalty weight is made to dominate the weight, so as to force the integral of the short-term load prediction sequence to strictly approximate the long-term load prediction value, thereby achieving a physical collaborative closed loop across time scales.

Images