Power demand evaluation method, system and device based on deep learning, and storage medium

Abstract

The invention discloses a power demand evaluation method, system and device based on deep learning, and a storage medium, and belongs to the technical field of power system intelligence and energy management, and the method comprises the steps: collecting data of multiple modes, and carrying out the preprocessing of the data; designing an adaptive weighting strategy according to the preprocessed data, weighting different modal data, constructing a deep learning model through weighted fusion data, and training the model; and outputting a prediction result of the power demand, and correcting the output result. According to the method, the dynamic weighting strategy of multi-modal data and a time window correction mechanism are fused, so that the adaptability of power demand prediction to complex environment changes is remarkably improved, and the load fluctuation characteristics of extreme weather can be accurately captured; and in combination with the time sequence modeling capability of the long-short-term memory network, the long-term dependency relationship and the multi-factor coupling rule are effectively learned, and the problem that a traditional model is insufficient in expression of periodic and sudden demand modes is solved.

Description

Technical Field

[0001] The present invention relates to the field of power system intelligence and energy management technology, and specifically to a method, system, device and storage medium for power demand assessment based on deep learning. Background Art

[0002] Electricity demand forecasting is a core task for ensuring stable power system operation and optimizing energy allocation. Its accuracy directly impacts power generation planning, grid dispatch efficiency, and power supply reliability. Traditional forecasting methods, such as time series analysis (ARIMA) and regression models, primarily rely on historical load data, inferring future trends through statistical patterns. However, with the transformation of energy structures and the diversification of electricity consumption scenarios, the factors influencing electricity demand are becoming increasingly complex. Frequent extreme weather events lead to short-term drastic load fluctuations, while factors such as holidays and economic activities introduce cyclical fluctuations. The integration of distributed energy resources further increases demand uncertainty. In this context, traditional methods have exposed a series of limitations. Relying solely on single-source data, historical load data makes it difficult to capture the coupled influence of multimodal information, such as meteorological conditions and holidays. This is particularly true during sudden heat waves or cold snaps, where forecast bias for fixed data increases significantly. Furthermore, they lack dynamic adaptability. Existing methods for fusing multi-source data typically use static weighting, failing to dynamically adjust the importance of each modal data based on real-time circumstances (such as typhoon warnings or unexpected holiday adjustments), resulting in inadequate representation of key features. Furthermore, time series modeling capabilities are limited. Traditional recurrent neural networks (RNNs) struggle to effectively capture long-term dependencies, while standard long-short-term memory models still face bottlenecks in their ability to learn cross-dimensional time series associations when processing multimodal data inputs in parallel. Finally, error correction mechanisms are lacking. Most methods directly output model predictions, lacking feedback to correct historical errors or real-time deviations. This problem of error accumulation is particularly prominent in medium- and long-term forecasts. For example, short-term load surges during cold waves can lead to forecast deviations for multiple consecutive days if not promptly corrected. While existing research has attempted to incorporate multi-source data or improve model structures, problems persist, such as the rigidity of meteorological data weights, inadequate cross-modal interaction modeling, and lagging correction strategies. Therefore, there is an urgent need for a power demand assessment method that can deeply integrate multimodal dynamic features, adaptively adjust data weights, enhance time series modeling capabilities, and implement closed-loop error correction. This approach can address forecasting challenges in complex scenarios and enhance the intelligent decision-making of power systems. Summary of the Invention

[0003] To solve the above technical problems, a method for power demand assessment based on deep learning is proposed, which includes collecting data from multiple modes and preprocessing the data;

[0004] Design an adaptive weighting strategy based on the preprocessed data, weight the data of different modalities, and generate a time series;

[0005] Build a long short-term memory network, use the time series as input data for training, and update the hidden state output prediction value;

[0006] Dynamically adjust the output results based on the prediction error to generate a revised prediction value;

[0007] The adaptive weighting strategy includes constructing a query vector of the current forecast demand and a key vector of historical modal features, calculating similarity and normalizing it to generate attention weights; imposing a time decay constraint so that the weights of historical data decay exponentially; defining the center and width parameters of the time window based on a Gaussian distribution, and dynamically adjusting the decay rate in combination with an adjustment coefficient; superimposing the attention weights and the time window correction weights, and weightedly fusing multimodal data to generate a time series input sequence.

[0008] As a preferred solution of the method for power demand assessment based on deep learning described in the present invention, wherein: the data collected in multiple modes is collected from various years of power load data, holiday information, and meteorological data;

[0009] The data preprocessing is to remove duplicate data, handle outliers, smooth short-term mutations in power load data, fill in missing parts using interpolation, standardize data of different modes, and use sliding window processing methods to extract long-term change trends from historical data.

[0010] As a preferred solution of the method for power demand assessment based on deep learning described in the present invention, wherein: the adaptive weighting strategy is designed based on the preprocessed data to weight different modal data, including:

[0011] The weight of each data point is dynamically calculated based on the current environment changes, so that the weight is adjusted as the input changes. The self-attention mechanism is introduced to calculate an importance weight for each modal data point. By calculating the similarity between the query vector and the key vector of the modal data point, the influence of different data points on the prediction target is determined.

[0012] By calculating the weight of the self-attention mechanism, a weight suitable for the current prediction task is assigned to each time step, and the weight is applied to the weighted sum of each data. The weighted data is the weighted fusion result of all modal data.

[0013] As a preferred embodiment of the method for power demand assessment based on deep learning described in the present invention, wherein: the method of designing an adaptive weighting strategy based on the preprocessed data, weighting different modal data, and generating a time series input sequence further includes designing a time window correction function;

[0014] Building a time window correction function involves designing a function that adaptively adjusts weights. This introduces a dynamic window correction method based on Gaussian distribution and correction coefficients, automatically adjusting the weighting strategy at each moment based on changes in time periods and fluctuations in demand.

[0015] The time window correction function is based on Gaussian distribution, with the center position and width of the window as parameters to adjust the weight of the data. The expression of the time window correction function is:

[0016]

[0017] Among them, Φ(t,δ) is the correction function, t represents the current position, σ represents the width parameter of the time window, controls the size of the window, λ represents the adjustment coefficient, μ δ Represented as the center position of the time window δ, σ δ It is represented as the width parameter of the time window, and λ is represented as the adjustment coefficient. It is expressed as the effect of adjusting the time window on the correction result;

[0018] Different weighting strategies are used in short-term and long-term forecasts, and a correction function based on a time window is constructed. The weighting strategies in different time periods are dynamically adjusted to adapt to different demand patterns. The expression of the weighted fusion time series is:

[0019]

[0020] in, is the weighted fusion time series, k is the total number of multimodal data, α i (t) is the importance weight of the i-th modal data at time step t, X i (t) is the eigenvalue of the i-th modal data at time step t.

[0021] As a preferred embodiment of the method for power demand assessment based on deep learning described in the present invention, the method comprises: constructing a long short-term memory network, training the time series as input data, and updating the hidden state output prediction value by using the adaptively weighted data as the input of the long short-term memory network, updating the state through the memory unit, and capturing the changing trend of the power load;

[0022] The weighted fusion time series is used as the time series input of the long short-term memory network. The long short-term memory network will gradually optimize its internal parameters based on the training data and output the predicted value of power demand.

[0023] The optimization of internal parameters is to optimize the trainable weights in the long short-term memory network, including the gate weight matrix, memory unit weights, and fully connected layer parameters.

[0024] As a preferred embodiment of the method for power demand assessment based on deep learning described in the present invention, the method of constructing a long short-term memory network, training the time series as input data, and updating the hidden state output prediction value includes generating a power demand prediction through the hidden state, and obtaining the power demand prediction value through linear transformation, which is expressed as follows:

[0025]

[0026] Among them, W y Represented as the weight matrix of the output layer, b y It is represented as the bias term of the output layer, Expressed as the predicted value of electricity demand, h t Represented as hidden state;

[0027] The mean square error is used as the loss function to optimize the power demand forecast value.

[0028] As a preferred embodiment of the method for power demand assessment based on deep learning according to the present invention, wherein: the output power demand prediction result and the correction of the output result include:

[0029] In electricity demand forecasting, the difference between actual electricity demand and forecast results is expressed as an error. A correction factor is introduced based on the error between actual demand and forecast to correct the forecast results. The corrected electricity forecast value is the sum of the electricity demand forecast value and the error caused by the introduced correction factor.

[0030] Another object of the present invention is to provide a system for power demand assessment based on deep learning, which solves the technical problems of insufficient multimodal data fusion, poor dynamic adaptability and accumulated prediction errors in traditional methods.

[0031] As a preferred solution of the system for power demand assessment based on deep learning described in the present invention, it is characterized by comprising: a multimodal data preprocessing module, an adaptive weighted fusion module, a long short-term memory deep learning modeling module, and a prediction result correction module;

[0032] The multimodal data preprocessing module collects multimodal data and performs data cleaning, smoothing and standardization, and uses sliding window technology to extract the long-term change trend of the data;

[0033] The adaptive weighted fusion module dynamically calculates the weights of each modal data, evaluates the importance of different data through the self-attention mechanism, and adjusts the weighting strategy according to the time period and demand fluctuations in combination with the time window correction function;

[0034] The long short-term memory deep learning modeling module inputs the weighted fusion data into the long short-term memory network, captures the temporal dependency through the memory unit, optimizes the model parameters using the mean square error as the loss function during the training process, and outputs the preliminary power demand forecast value;

[0035] The prediction result correction module analyzes the error between the predicted value and the actual demand, introduces a correction factor to correct the prediction result, and optimizes the final prediction accuracy.

[0036] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method for power demand assessment based on deep learning when executing the computer program.

[0037] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for power demand assessment based on deep learning.

[0038] The beneficial effects of the present invention are as follows: by integrating the dynamic weighting strategy and time window correction mechanism of multimodal data, the adaptability of power demand forecasting to complex environmental changes is significantly improved, and the load fluctuation characteristics in extreme weather, holidays and other scenarios can be accurately captured; combined with the time series modeling capability of the long short-term memory network, it effectively learns long-term dependencies and multi-factor coupling laws, and solves the problem that traditional models are insufficient in expressing periodic and sudden demand patterns; further, through the closed-loop error correction module, the prediction deviation is compensated in real time, and the error accumulation is suppressed, ultimately achieving highly robust, multi-scenario compatible power demand assessment, providing a reliable basis for power system optimization and scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 An overall flow chart of a method for power demand assessment based on deep learning provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0041] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0042] Example 1, reference Figure 1 , which is the first embodiment of the present invention, provides a method for power demand assessment based on deep learning, including:

[0043] Collect data of multiple modalities and preprocess the data.

[0044] An adaptive weighting strategy is designed based on the preprocessed data to weight different modal data and generate time series.

[0045] Build a long short-term memory network, use the time series as input data for training, and update the hidden state output prediction value.

[0046] The output results are dynamically adjusted based on the prediction error to generate a revised prediction value.

[0047] Collect historical power load data by connecting to the data interface of power companies or the State Grid. Extract historical load information on demand through API or database connection.

[0048] By establishing cooperation with meteorological data providers, we can obtain the corresponding meteorological data interface and obtain it by hour or day.

[0049] Obtain holiday information from the public holiday database or government website and connect the holiday information with historical power load data.

[0050] The collected data is first processed to remove duplicate data and address outliers. Short-term sudden changes in power load data are smoothed, and missing data are filled using interpolation. Data of different modalities are standardized to minimize the impact of varying data ranges on subsequent model training. Furthermore, a sliding window approach is used to extract long-term trends from historical data, enabling better feature extraction.

[0051] During the multimodal data fusion process, we collected input data from multiple sources, including historical power load data, meteorological data, holiday information, and socioeconomic data. Each data point has a varying degree of influence on power demand forecasts. Therefore, dynamically adjusting the weight of each data point based on time and environmental changes is key to improving forecast accuracy.

[0052] The core of the adaptive weighting strategy is to dynamically calculate the weight of each data point based on the current environment, so that the weights adjust as the input changes. The main advantage of this strategy is that it can automatically learn and adapt to fluctuations in external conditions, avoiding the limitations of manually set static weights.

[0053] The self-attention mechanism is introduced in the adaptive weighting strategy. This mechanism can calculate an importance weight for each modal data. By calculating the similarity between the query vector and the key vector of the modal data, the influence of different data on the prediction target is determined.

[0054] Specifically, a query vector and a corresponding key vector are first generated for each modal data. The query vector and the corresponding key vector jointly determine the temporal weight of the modal data.

[0055] By calculating the weights of the self-attention mechanism, we can assign a weight suitable for the current prediction task to each time step and apply these weights to the weighted sum of each data. The weighted data is the weighted fusion result of all modal data.

[0056] The weight calculation formula is as follows:

[0057]

[0058] Among them, Q t Denoted as query vector, K i,t Represented as the key vector of mode i at time t, α i,t Expressed as the contribution of mode i to the current power demand forecast, Q t is represented as the query vector of the modality at time t.

[0059] The weighted data fusion results are as follows:

[0060]

[0061] in, It is expressed as the data fusion result after adaptive weighting, α i,t It is represented as the modified weight of each modality i at time t.

[0062] Given the temporal nature of electricity demand, demand patterns can vary significantly across time periods. By flexibly adjusting the size and center position of the time window, we can adapt to cyclical fluctuations such as seasonality and weekend effects. By analyzing historical data trends, we dynamically adjust weights to account for long-term trends, ensuring that the weights appropriately reflect seasonal and cyclical fluctuations in electricity demand. The time window correction function dynamically adjusts the weight of each data point based on factors such as the distance between the current and past moments and demand fluctuations, enabling the model to adapt adaptively.

[0063] The key step in building a time window correction function is designing a function that adaptively adjusts weights. We introduced a dynamic window correction method based on Gaussian distribution and correction coefficients, which automatically adjusts the weighting strategy at each moment based on changes in time periods and fluctuations in demand.

[0064] First, the time window correction function is based on the Gaussian distribution, with the center position and width of the window as parameters, and the weight of the data is adjusted according to these parameters. The Gaussian distribution has a strong smoothness, which can ensure that the weight of the moments close to the center of the time window is higher, while the weight of the moments away from the center gradually decreases. The designed time window correction function is specifically expressed as:

[0065]

[0066] Among them, Φ(t,δ) is the correction function, t represents the current position, μ represents the center position of the time window, σ represents the width parameter of the time window, which controls the size of the window, and λ represents the adjustment coefficient.

[0067] The numerator of this function represents the core of the Gaussian function, describing the effect of the distance from time t to the center of the time window μ on the weight. The closer the distance, the larger the exponent value and the corresponding higher the weighting coefficient. The denominator of this function acts as a correction factor, adjusting the weight's influence in certain circumstances. By adjusting the values ​​of λ and σ, the influence of different time periods on the results can be controlled.

[0068] In order to further improve the accuracy and flexibility of weight adjustment, the concept of time window is introduced to allow different weighting strategies to be used in short-term and long-term predictions. Specifically, a correction function based on time series window is designed. This correction function is used to dynamically adjust the weighting strategy in different time periods to adapt to different demand patterns. The expression of weighted fusion time series is:

[0069]

[0070] in, is the weighted fusion time series, μ δ Represented as the center position of the time window δ, σδ It is represented as the width parameter of the time window, and λ is represented as the adjustment coefficient. It represents the influence of adjusting the time window on the correction result, and Φ(t,δ) is represented as the correction function.

[0071] By dynamically adjusting the center and width of the window, the time window correction function can flexibly respond to changes in electricity demand in different time periods. Especially in predictions for special periods such as seasonality and holidays, it can significantly improve the prediction accuracy of the model.

[0072] The introduction of Gaussian distribution makes the adjustment of weights smoother and avoids excessive impact of sudden changes on prediction results.

[0073] Combined with the trend correction term, the model can better capture the patterns of long-term demand fluctuations, thereby improving the accuracy of long-term forecasts.

[0074] When forecasting electricity demand, in addition to integrating multimodal data, deep learning models, particularly time series models, can better capture the temporal dependencies and long-term trends of electricity demand. Long Short-Term Memory (LSTM) networks, when processing time series data, effectively address the challenges of traditional RNNs (recurrent neural networks) in learning long-term dependencies.

[0075] Long short-term memory can capture the potential patterns in historical data and weighted fused data in the time dimension, thereby outputting accurate power demand forecasts.

[0076] The weighted fused time series is fed into the LSTM network, which processes this time series data, gradually learning the dependencies between different time steps to generate a forecast of future electricity demand.

[0077] Specifically, the long-short-term memory network updates its state through multiple memory cells to capture the changing trends of power load. The input weighted data will serve as the time series input of the model, and the network will gradually optimize its internal parameters based on the training data to achieve the best prediction results.

[0078] It should be further explained that the internal parameters to be optimized are the trainable weights of the long short-term memory network, including the gate weight matrix: which controls the calculation rules of the input gate, forget gate, and output gate; the memory unit weight: which determines the intensity of retaining historical information and writing new information; and the fully connected layer parameters: which map the hidden state to the weight and bias of the final prediction value.

[0079] All long short-term memory gating weight matrices and fully connected layer parameters are initialized using Xavier normal distribution, and the bias term is initialized to 0 to ensure training stability.

[0080] Input the weighted fusion time series into the long short-term memory network and perform the following operations in sequence according to the time step:

[0081] Forget gate: Calculates the retention ratio of the memory unit at the previous moment.

[0082] Input gate: Calculates the update strength of the current input data on the memory unit.

[0083] Memory unit update: Update the memory unit state according to the results of the forget gate and the input gate.

[0084] Output gate: Generates hidden state output based on the current memory unit.

[0085] The mean squared error (MSE) is calculated for the predicted value and the true value at each time step, and an L2 regularization term (0.01 times the sum of squares of all parameters) is added to the total loss to prevent the model from overfitting.

[0086] The Adam optimization algorithm is used to automatically adjust the learning rate and update the weights according to the gradient direction of the loss function to the parameters:

[0087] Calculate gradients: Backpropagate the error signal through the chain rule to obtain the gradient values ​​of each parameter.

[0088] Parameter update: Adjust the parameter value according to the gradient direction. The initial learning rate is 0.001. If the validation loss does not decrease for 5 consecutive rounds, the learning rate decays to 50% of the current value.

[0089] Training is terminated when any of the following conditions are met:

[0090] The validation set loss has not decreased for 10 consecutive rounds; the total number of training rounds has reached the preset upper limit (100 rounds).

[0091] The weighted fusion time series is used as the input data of the long short-term memory network. The weighted data sequence is specifically expressed as follows.

[0092]

[0093] Among them, X LSTM Represented as a weighted modal data sequence from time t-δ to time t.

[0094] Long Short-Term Memory Network Design:

[0095] The LSTM network design consists of multiple stacked LSTM units. Each LSTM layer controls the flow of information through internal forget gates, input gates, and output gates, enabling the memorization and updating of information based on long-term dependencies. Sigmoid activation functions control the gating mechanism, while hyperbolic tangent activation functions perform nonlinear transformations on the input. These gating mechanisms gradually update the state and memory cells of the LSTM units, passing them to the next layer or as output.

[0096] It is further explained that: the input sequence X LSTM Expanded by time step: Each time step corresponds to a weighted data point.

[0097] The long short-term memory network processes the input of each time step in turn, combines the hidden state and memory unit of the previous moment through the gating mechanism (forget gate, input gate, output gate), and updates the current hidden state and memory unit one by one.

[0098] When the entire input sequence X is processed LSTM After all time steps, the final hidden state It will contain the temporal feature information within the entire time window.

[0099] The final hidden state Perform linear transformation (Formula ), output power demand forecast value

[0100] The final output of the LSTM network is a predicted value for electricity demand. The LSTM generates a demand forecast using its hidden state. We perform a linear transformation on the hidden state to obtain the predicted demand value.

[0101]

[0102] Among them, W y Represented as the weight matrix of the output layer, b y It is represented as the bias term of the output layer, Expressed as the predicted value of electricity demand, h t Represented as hidden state.

[0103] The mean square error is used as the loss function to optimize the model. The loss function is as follows.

[0104]

[0105] Among them, y t Expressed as actual power demand, It represents the predicted power demand and T represents the sample size.

[0106] During training, we use the Adam optimizer to update parameters. The Adam optimizer combines the advantages of gradient descent and adaptive learning rates to more effectively optimize the parameters of the LSTM model. The specific formula for the Adam optimizer is as follows.

[0107]

[0108] Among them, α t Denoted as the learning rate, Expressed as the first moment estimate of the gradient, It is expressed as the second-order moment estimate of the gradient, ∈ represents a small constant,

[0109] When processing time series data, the Long Short-Term Memory (LSTM) network can capture trend fluctuations in electricity demand and flexibly adapt to factors such as seasonality and holidays. Furthermore, through a correction mechanism, we ensure that electricity demand forecasts are more accurate, especially in complex and changing electricity demand scenarios.

[0110] Correcting output results. In power demand forecasting tasks, the model's output is typically a predicted value for power demand. However, these predictions can be affected by data noise, historical data bias, and external factors not considered. Therefore, after obtaining a preliminary forecast, we need to correct the output results to improve the accuracy and reliability of the forecast.

[0111] The purpose of correcting forecast results is to reduce errors in model predictions and improve their accuracy through real-time feedback. In power demand forecasting, the difference between actual power demand and the forecast is expressed as error. The specific formula for calculating error is as follows.

[0112]

[0113] Among them, y t Expressed as the actual power demand at time t, Denotes the predicted power demand obtained by the deep learning model, Δy t,corr Expressed as error.

[0114] In order to correct the prediction results, a correction factor is introduced. Its size is determined by the error. According to actual needs and the error predicted by the model, the correction factor calculation formula is as follows.

[0115] Δy t,corr =η·(Δy t )

[0116] Where Δy t,corr It is represented as the correction factor, and η is represented as the learning rate of the correction factor.

[0117] The specific calculation formula for the revised power forecast value is as follows.

[0118]

[0119] in, Expressed as the revised electricity demand forecast value, Represents the preliminary electricity demand predicted by the long short-term memory model.

[0120] Example 2 is the second embodiment of the present invention, which provides a method for power demand assessment based on deep learning. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0121] The data range is from January 1, 2020 to December 31, 2023, covering three full years, including weekdays, weekends, statutory holidays and extreme weather events (such as cold waves and high temperatures).

[0122] Power load data: Actual load values ​​(unit: MW) recorded every hour for a provincial power grid, totaling 35,064 data points.

[0123] Meteorological data: hourly temperature (°C), relative humidity (%), and wind speed (m / s), obtained from the Meteorological Bureau API and time-aligned with the load data.

[0124] Holiday information: National statutory holidays and their corresponding dates, marked as binary variables (1 for holidays and 0 for non-holidays).

[0125] ARIMA model: a classic time series model that uses only historical load data, with parameters (p=3, d=1, q=2) optimized using the AIC criterion.

[0126] Static weighted long and short-term memory: fixed weight fusion of multimodal data (meteorological weight 10%, holiday weight 5%, load history data 85%), 64 long and short-term memory hidden layer units.

[0127] The method of the present invention: dynamic weighted long short-term memory (self-attention mechanism + time window correction) + closed-loop error correction module, 64 hidden layer units, time window parameter σ = 24 (hours), Gaussian correction function λ = 0.5, σ δ =6.

[0128] MAE and REMS are used as evaluation indicators.

[0129] Table 1 Comparison of prediction accuracy

[0130]

[0131] During the cold wave, the proposed method achieved a MAE of 148.7MW, a 29.9% reduction compared to the static weighted LSTM (212.4MW). This is due to the self-attention mechanism detecting a sudden drop in temperature (-8°C) in the first hour of the cold wave and dynamically increasing the weight of meteorological data from an initial 10% to 65% within 30 minutes, thus strengthening the modeling of the correlation between low temperatures and heating load.

[0132] In high-temperature scenarios, the RMSE dropped from 180.3 MW to 135.7 MW, a 24.7% reduction. The time window correction function increases the meteorological weight by an additional 20% during the high-temperature period (12:00-18:00), capturing the sharp increase in air conditioning load in the afternoon.

[0133] The static weighted long-term and short-term memory underestimates the load increase before the holidays due to the fixed holiday weight (5%), resulting in a MAE as high as 162.5MW. δ Set it to 3 days before New Year's Eve, σ δ =5 days), and gradually increased the holiday weight to 50% in the week before the holiday, accurately predicting the load increase during the peak period of returning home.

[0134] Table 2 Dynamic adaptability analysis (key modal weight adjustment)

[0135]

[0136] As shown in Table 2, after detecting a sudden change in temperature, the self-attention mechanism completes the weight adjustment (10% → 65%) within 30 minutes. The technical principle is to match the similarity between the query vector (current load mutation feature) and the meteorological key vector (temperature, wind speed) in real time.

[0137] Static weighted long short-term memory cannot respond to sudden meteorological events due to its fixed weights. The forecast error of the first day of the cold wave is as high as 280MW, while the error of the present invention is only 150MW.

[0138] The time window correction function increases the holiday weight by an average of 6.4% per day over the seven days leading up to the holiday, ultimately reaching 50% on the day before the holiday. The actual effect: The load forecast error in the three days leading up to the holiday is reduced from 15% in the static model to 5%.

[0139] Table 3 Error suppression effect

[0140]

[0141]

[0142] The static model forecast on the third day of the cold wave was 200MW lower. The present invention compensated for 120MW through the correction factor in the forecast on the fourth day, reducing the error to 80MW.

[0143] Static weighted long-term short-term memory (SWTSM) cannot correct for continuous forecast deviations, resulting in a 7-day cumulative error of 1512MW. The present invention uses a closed-loop correction module to calculate a correction factor daily based on the previous day's error, suppressing the cumulative error to 905MW, a suppression rate of 38.2%.

[0144] In complex scenarios such as cold waves, high temperatures, and the Spring Festival, the system reduces the average prediction error (MAE) by more than 30% compared with traditional methods through dynamic weighted fusion and time window correction. During the cold wave, the error dropped sharply from 212.4MW to 148.7MW (a decrease of 29.9%), breaking through the bottleneck of static models in responding to emergencies.

[0145] The self-attention mechanism enables rapid adjustment of key modal weights (for example, the meteorological weight increases to 65% within 30 minutes during a cold wave), and the time window correction function supports long-term progressive adaptation (for example, the weight of holidays before festivals increases linearly to 50% for 7 days). The two work together to solve the rigidity of traditional methods in timing and scene adaptation.

[0146] From deep fusion of multimodal data, dynamic weighted strategy optimization, to long- and short-term memory time series modeling and closed-loop correction, the present invention forms a complete technical chain of "data-driven → model iteration → feedback optimization", providing high-precision, high-robustness, and multi-scenario compatible decision support for power system dispatching.

[0147] Example 3 is the third embodiment of the present invention, which differs from the first two embodiments in that:

[0148] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0149] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0150] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0151] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0152] Example 4 is the fourth embodiment of the present invention. This embodiment provides a system for power demand assessment based on deep learning, including a multimodal data preprocessing module, an adaptive weighted fusion module, a long and short-term memory deep learning modeling module, and a prediction result correction module.

[0153] The multimodal data preprocessing module collects multimodal data and performs data cleaning, smoothing and standardization, and uses sliding window technology to extract the long-term change trend of the data.

[0154] The adaptive weighted fusion module dynamically calculates the weights of each modal data, evaluates the importance of different data through the self-attention mechanism, and adjusts the weighting strategy according to the time period and demand fluctuations in combination with the time window correction function.

[0155] The long short-term memory deep learning modeling module inputs the weighted fusion data into the long short-term memory network, captures the temporal dependency through the memory unit, optimizes the model parameters using the mean square error as the loss function during the training process, and outputs the preliminary power demand forecast value.

[0156] The prediction result correction module analyzes the error between the predicted value and the actual demand, introduces correction factors to correct the prediction results, and optimizes the final prediction accuracy.

[0157] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for power demand assessment based on deep learning, characterized by: include, Collect data from multiple modalities and pre-process the data; Design an adaptive weighting strategy based on the preprocessed data, weight the data of different modalities, and generate a time series; Build a long short-term memory network, use the time series as input data for training, and update the hidden state output prediction value; Dynamically adjust the output results based on the prediction error to generate a revised prediction value; The adaptive weighting strategy includes constructing a query vector of the current forecast demand and a key vector of historical modal features, calculating similarity and normalizing it to generate attention weights; Applying time decay constraints makes the weight of historical data decay exponentially; The center and width parameters of the time window are defined based on the Gaussian distribution, and the attenuation rate is dynamically adjusted in combination with the adjustment coefficient; the attention weight is superimposed with the time window correction weight, and the multimodal data is weightedly fused to generate a time series input sequence.

2. The method for power demand assessment based on deep learning according to claim 1, wherein: The data collected in multiple modes is to collect power load data, holiday information, and meteorological data over the years; The data preprocessing is to remove duplicate data, handle outliers, smooth short-term mutations in power load data, fill in missing parts using interpolation, standardize data of different modes, and use sliding window processing methods to extract long-term change trends from historical data.

3. The method for power demand assessment based on deep learning according to claim 2, wherein: The method of designing an adaptive weighting strategy based on the preprocessed data, weighting the different modal data, and generating a time series input sequence includes: The weight of each data point is dynamically calculated based on the current environment changes, so that the weight is adjusted as the input changes. The self-attention mechanism is introduced to calculate an importance weight for each modal data point. The self-attention mechanism is introduced to construct a query vector and a key vector for each modal data, where the query vector represents the predicted target demand at the current moment, and the key vector represents the modal characteristics of the historical period; Calculate the similarity between the query vector and the key vector. The similarity is proportional to the influence weight of the corresponding modality on the current prediction. Normalize the similarity to obtain the attention weight distribution of each modality in the time window; Apply a time decay constraint to the attention weight, so that the weight of historical data farther away from the current moment decays exponentially; The attention weight is multiplied by the time decay coefficient to generate a final time-space sensitive weight, and the weighted sum of each modal data is performed. Taking the current moment as the benchmark, the weighted results of multiple consecutive moments are spliced ​​into a time series input sequence according to the preset time window length.

4. The method for power demand assessment based on deep learning according to claim 3, wherein: The method of designing an adaptive weighting strategy based on the preprocessed data, weighting the different modal data, and generating a time series input sequence also includes designing a time window correction function; Building a time window correction function involves designing a function that adaptively adjusts weights. This introduces a dynamic window correction method based on Gaussian distribution and correction coefficients, automatically adjusting the weighting strategy at each moment based on changes in time periods and fluctuations in demand. The time window correction function is based on Gaussian distribution, with the center position and width of the window as parameters to adjust the weight of the data. The expression of the time window correction function is: Among them, Φ(t,δ) is the correction function, t represents the current position, σ represents the width parameter of the time window, controls the size of the window, λ represents the adjustment coefficient, μ δ Represented as the center position of the time window δ, σ δ It is represented as the width parameter of the time window, and λ is represented as the adjustment coefficient. It is expressed as the effect of adjusting the time window on the correction result; Different weighting strategies are used in short-term and long-term forecasts, and a correction function based on a time window is constructed. The weighting strategies in different time periods are dynamically adjusted to adapt to different demand patterns. The expression of the weighted fusion time series is: in, is the weighted fusion time series, k is the total number of multimodal data, α i (t) is the importance weight of the i-th modal data at time step t, X i (t) is the eigenvalue of the i-th modal data at time step t.

5. The method for power demand assessment based on deep learning according to claim 4, characterized in that: The long short-term memory network is constructed, and the time series is used as input data for training. The hidden state output prediction value is updated by using the adaptively weighted data as the input of the long short-term memory network, updating the state through the memory unit, and capturing the changing trend of the power load; The weighted fusion time series is used as the time series input of the long short-term memory network. The long short-term memory network will gradually optimize its internal parameters based on the training data and output the predicted value of power demand. The optimization of internal parameters is to optimize the trainable weights in the long short-term memory network, including the gate weight matrix, memory unit weights, and fully connected layer parameters.

6. The method for power demand assessment based on deep learning according to claim 5, characterized in that: The long short-term memory network is constructed, the time series is used as input data for training, and the hidden state output prediction value is updated. The prediction of power demand is generated by the hidden state, and the power demand prediction value is obtained through linear transformation. The expression is: Among them, W y Represented as the weight matrix of the output layer, b y Represented as the bias term of the output layer, Expressed as the predicted value of electricity demand, h t Represented as hidden state; The mean square error is used as the loss function to optimize the power demand forecast value.

7. The method for power demand assessment based on deep learning according to claim 6, wherein: The output power demand prediction result, and the correction of the output result include: In electricity demand forecasting, the difference between actual electricity demand and forecast results is expressed as an error. A correction factor is introduced based on the error between actual demand and forecast to correct the forecast results. The corrected electricity forecast value is the sum of the electricity demand forecast value and the error caused by the introduced correction factor.

8. A system for power demand assessment based on deep learning, applying the method for power demand assessment based on deep learning according to any one of claims 1 to 7, characterized in that: Including: multimodal data preprocessing module, adaptive weighted fusion module, long-term and short-term memory deep learning modeling module, and prediction result correction module; The multimodal data preprocessing module collects multimodal data and performs data cleaning, smoothing and standardization, and uses sliding window technology to extract the long-term change trend of the data; The adaptive weighted fusion module dynamically calculates the weights of each modal data, evaluates the importance of different data through the self-attention mechanism, and adjusts the weighting strategy according to the time period and demand fluctuations in combination with the time window correction function; The long short-term memory deep learning modeling module inputs the weighted fusion data into the long short-term memory network, captures the temporal dependency through the memory unit, optimizes the model parameters using the mean square error as the loss function during the training process, and outputs the preliminary power demand forecast value; The prediction result correction module analyzes the error between the prediction value and the actual demand, introduces a correction factor to correct the prediction result, and optimizes the final prediction accuracy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for power demand assessment based on deep learning according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for power demand assessment based on deep learning according to any one of claims 1 to 7 are implemented.

Images