A multi-type power supply day-ahead demand prediction system for dispatch operation and maintenance

Abstract

The application belongs to the technical field of demand prediction, and specifically relates to a multi-type power source day-ahead demand prediction system for dispatching operation and maintenance, which comprises a heterogeneous data processing module for generating time sequence characteristic data sets. A multi-type prediction module outputs day-ahead prediction curves of various power sources. A coupling and collaborative correction module generates a collaborative prediction scenario set with a confidence interval. A dispatching constraint adaptation module embeds the collaborative prediction scenario set into a power grid dispatching hard constraint, performs feasibility verification and boundary correction, and generates multi-type power source demand prediction results. A rolling closed-loop iteration module is used for online updating of each prediction unit to generate an iteration-optimized model parameter set. The application can guarantee the integrity and consistency of input data, improve the perception ability and prediction accuracy for complex operating environments, break the isolated island of independent prediction of various power sources, and improve the prediction rationality of the overall system through collaborative correction.

Description

Technical Field

[0001] This invention belongs to the field of demand forecasting technology, specifically a day-ahead demand forecasting system for various power sources oriented towards scheduling and maintenance. Background Technology

[0002] With the continuous advancement of the construction of new power systems, a high proportion of new energy sources, diversified energy storage, and flexible loads are widely connected to the power grid, significantly increasing the randomness and volatility on both the source and load sides. As a core supporting technology for dispatching and operation and maintenance, day-ahead demand forecasting, its forecasting accuracy and engineering feasibility directly determine the execution quality of day-ahead unit combination, economic dispatch, and safety verification.

[0003] Current day-ahead forecasting technologies for various power sources still suffer from several shortcomings. First, the foundation of heterogeneous multi-source data is weak. Data from thermal power, hydropower, wind and solar power plants, energy storage, and loads are collected by different systems, resulting in significant differences in sampling frequency, time-series accuracy, and data quality. Frequent missing outliers and insufficient cross-source time-series alignment lead to insufficient reliability of the forecasting model input, fundamentally limiting the upper limit of forecast accuracy. Second, model adaptability is insufficient. Existing solutions often use a unified architecture to adapt to multiple power sources, failing to consider the physical operational constraints of controllable power sources, the meteorological randomness of new energy sources, the bidirectional regulation characteristics of energy storage, and the cyclical variation patterns of loads. This makes it impossible to match the different output characteristics of various power sources, leading to a significant increase in forecast errors under fluctuating operating conditions. Third, forecasting and dispatching operations are disconnected. Existing technologies often use independent modeling modes for each power source, ignoring spatiotemporal coupling relationships such as wind-solar complementarity and source-load synchronization. System-level net load errors are easily accumulated and amplified, making it difficult to directly support day-ahead dispatching and maintenance decisions.

[0004] Therefore, there is an urgent need to develop a multi-source day-ahead demand forecasting system for dispatching and maintenance, in order to meet the dispatching and maintenance needs of the new power system. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a multi-source day-ahead demand forecasting system for scheduling and operation. This invention primarily addresses the problems in traditional forecasting systems, such as poor quality of heterogeneous multi-source data, significant differences in output characteristics among various power sources, and a lack of coordination, which lead to forecast results deviating from actual grid constraints.

[0006] The technical solution adopted by this invention to solve its technical problem is: This invention provides a multi-type power day-ahead demand forecasting system for scheduling and maintenance, comprising: The heterogeneous data processing module is used to collect various types of power source operation data, power load data, weather forecast data, and scheduling constraint data, and to preprocess them to generate time-series feature datasets.

[0007] The multi-type prediction module is used to read time-series feature datasets, classify them according to power supply operating characteristics, construct prediction units for each category, and output day-ahead prediction curves for each type of power supply in parallel.

[0008] The coupled collaborative correction module is used to mine spatiotemporal coupling correlation patterns based on the day-ahead prediction curves of various power sources, perform multivariate joint correction and uncertainty quantification, and generate a set of collaborative prediction scenarios with confidence intervals.

[0009] The scheduling constraint adaptation module is used to embed the collaborative prediction scenario set into the hard constraints of power grid scheduling, perform feasibility verification and boundary correction, and generate multiple types of power demand prediction results.

[0010] The rolling closed-loop iteration module is used to collect actual intraday operating data, calculate prediction deviations and identify the degree of data distribution drift, update the execution of each prediction unit online, and generate the iteratively optimized model parameter set.

[0011] This invention provides a multi-type power day-ahead demand forecasting system for scheduling and maintenance. The heterogeneous data processing module performs the following preprocessing steps: The system collects operation sequence data of thermal power and hydropower, output data of wind power and photovoltaic power stations, operation data of energy storage charging and discharging, load data of the entire network and its zones, grid data of numerical weather forecasts, and maintenance data of generating units to form a raw multi-source dataset.

[0012] Outlier identification and removal are performed on the original multi-source dataset, missing data are repaired using nearest neighbor interpolation, and the scale of the entire dataset is normalized to obtain the basic dataset.

[0013] The basic dataset is timestamped using a preset standard time step to unify the data sampling frequency and obtain a time-aligned dataset.

[0014] This invention provides a multi-type power day-ahead demand forecasting system for scheduling and maintenance. The heterogeneous data processing module generates a time-series feature dataset in the following steps: Basic operational features are extracted from the time-series aligned dataset. Calendar features are constructed by combining date type, seasonal attributes, and holiday identifiers. Meteorological driving features are constructed by combining temperature, irradiance, wind speed, and cloud cover. Physical constraint features are constructed by combining unit rated parameters and grid structure parameters, forming a multi-dimensional feature factor library.

[0015] The multi-dimensional feature factor library is mapped one-to-one with the power output labels and load demand labels of the corresponding time period according to the timestamp, and a labeled feature sample set is constructed.

[0016] A rolling time window mechanism is used to divide the feature sample set into training, validation and test sets to generate a time series feature dataset.

[0017] This invention provides a multi-type power supply day-ahead demand forecasting system for scheduling and maintenance. The steps of the multi-type forecasting module outputting day-ahead forecast curves for various power supplies include: The time-series feature dataset is divided into controllable power source subset, fluctuating new energy subset, energy storage operation subset, and power load subset according to power source controllability and output characteristics.

[0018] For a subset of controllable power sources, a time-series prediction model is constructed by combining the unit ramp rate and minimum output physical constraints, and the adjustable output baseline curves of thermal power and hydropower are output.

[0019] For a subset of fluctuating renewable energy sources, a decomposition-integration deep learning framework is adopted to construct an output prediction model by coupling numerical weather forecast features, and output day-ahead output prediction curves for wind power and photovoltaic power.

[0020] For a subset of energy storage operations, a boundary calculation model is constructed by combining initial state of charge, charge / discharge efficiency, and lifetime decay constraints, and the boundary curve of the chargeable / dischargeable power of the energy storage is output.

[0021] For a subset of electricity load, a load forecasting model is constructed by combining calendar features and meteorological sensitivity factors, and the day-ahead demand curve of regional electricity load is output.

[0022] The adjustable output baseline curve, day-ahead output forecast curve, chargeable and dischargeable power boundary curve, and regional power load day-ahead demand curve are spliced ​​and integrated according to a unified time step to form day-ahead forecast curves for various power sources.

[0023] The present invention provides a multi-type power source day-ahead demand forecasting system for scheduling and operation and maintenance, which outputs day-ahead output forecast curves for wind power and photovoltaic power, including the following steps: The ensemble empirical mode decomposition algorithm is used to decompose the historical output sequence of new energy into low-frequency trend components that characterize long-term changes and high-frequency random components that characterize short-term fluctuations.

[0024] The low-frequency trend component and the trend meteorological characteristics in numerical weather prediction are input into a temporal convolutional network to fit the long-term variation law of new energy output and generate low-frequency prediction components.

[0025] The high-frequency random components and spatial correlation features of adjacent stations are input into a graph convolutional network to capture the wave transmission and spatial complementarity characteristics between stations in the region, and generate high-frequency prediction components.

[0026] By superimposing and reconstructing the low-frequency and high-frequency prediction components, the day-ahead power output prediction curves for wind power and photovoltaic power are obtained.

[0027] This invention provides a multi-type power day-ahead demand forecasting system for scheduling and maintenance, the coupled collaborative correction module including: The coupling correlation mining unit is used to read the day-ahead forecast curves of various power sources, extract the coupling features of wind-solar complementarity, source-load synchronization, and hydro-thermal substitution, and construct a multivariate correlation correction matrix.

[0028] The collaborative baseline correction unit is used to perform spatiotemporal joint correction on the daily forecast curves through a multivariate correlation correction matrix, and generate a collaborative correction baseline curve.

[0029] The uncertainty quantification unit is used to statistically analyze the distribution characteristics of historical prediction errors, fit the error probability distribution function, and calculate the upper and lower bounds of the prediction interval at different confidence levels.

[0030] The typical scenario generation unit is used to generate an initial set of running scenarios based on the prediction interval, and after scenario reduction, typical running scenarios are extracted to generate a collaborative prediction scenario set.

[0031] This invention provides a multi-type power day-ahead demand forecasting system for scheduling and maintenance, wherein a typical scenario generation unit includes: The scene sampling unit is used to generate an initial set of running scenes based on the error probability distribution and prediction interval using the Latin hypercube sampling method. Each scene corresponds to a set of source load time series curves.

[0032] The scenario reduction subunit is used to cluster and reduce the initial set of running scenarios using the K-means clustering algorithm, retaining high-probability regular scenarios and low-probability extreme risk scenarios to obtain typical running scenarios.

[0033] The scenario integration subunit is used to integrate and encapsulate typical operating scenarios with collaborative correction benchmark curves and upper and lower bounds of confidence intervals to generate a collaborative prediction scenario set.

[0034] This invention provides a multi-type power source day-ahead demand forecasting system for scheduling and operation, wherein the steps of the scheduling constraint adaptation module in generating multi-type power source demand forecasting results include: The baseline curve and time-period output and load values ​​of each typical scenario are extracted from the collaborative prediction scenario set.

[0035] The system incorporates unit operation constraints such as unit ramp rate, minimum technical output, and minimum start-up / shutdown time, as well as grid security constraints such as cross-sectional transmission limits and spinning reserve ratio, to construct a scheduling verification rule base.

[0036] Based on the scheduling verification rule base, the feasibility of the collaborative correction benchmark curve is verified on a time-by-time basis. For time periods exceeding the limit, the output boundary is smoothed and corrected to generate a benchmark prediction curve that meets the scheduling hard constraints.

[0037] In accordance with the scheduling business specifications, the baseline forecast curve, confidence interval and typical scenarios are output in a preset time sequence format to generate multi-type power demand forecast results that can be directly used for day-ahead scheduling, categorized by region and power type.

[0038] This invention provides a multi-type power day-ahead demand forecasting system for scheduling and maintenance. The rolling closed-loop iterative module identifies the degree of data distribution drift in the following steps: The actual daily operation data is collected and compared point by point with the day-ahead forecast results for the corresponding time period to calculate the forecast deviation sequence by power source type and time period.

[0039] Based on the prediction bias sequence, the statistical error distribution characteristics of the sliding window are used, and the concept drift detection algorithm is used to determine the degree of deviation between the current data distribution and the training data distribution, and generate drift degree labels.

[0040] When the drift exceeds a preset threshold, an incremental update instruction is triggered to extract the latest time period's running data and construct an incremental training set.

[0041] This invention provides a multi-type power day-ahead demand forecasting system for scheduling and maintenance, comprising: The steps of the rolling closed-loop iteration module in generating the model parameter set include: The incremental training set is read, and the model parameters of each prediction unit are fine-tuned and optimized by using transfer learning and incremental training.

[0042] The weight coefficients of each driving feature in the feature factor library are updated synchronously, and the feature importance ranking is corrected.

[0043] The optimized model parameters and feature weights are solidified and encapsulated to generate a model parameter set, which is then distributed to each prediction unit to complete the model iteration.

[0044] The beneficial effects of this invention are as follows: This invention eliminates differences in format, frequency, and units among multi-source data by processing heterogeneous data and constructing multi-dimensional features, ensuring the integrity and consistency of the input data. Deeply integrating physical mechanisms with data-driven features significantly improves the model's perception capability and prediction accuracy in complex operating environments. It effectively decouples the long-term trend and short-term random fluctuations of renewable energy output, while utilizing the spatial complementarity between power plants within the region, significantly improving the prediction accuracy of wind and solar day-ahead output. It breaks down the isolated predictions of various power sources, improving the overall system's prediction rationality through collaborative correction. The generated prediction results include confidence intervals and typical operating scenarios, fully quantifying the uncertainty of the prediction. It can keenly capture data distribution drift caused by seasonal changes, extreme weather, or changes in grid topology. Through online iterative optimization, it ensures high stability and high accuracy of the prediction model during long-term operation, extending the model's lifespan. Attached Figure Description

[0045] The invention will now be further described with reference to the accompanying drawings.

[0046] Figure 1 This is a schematic diagram of a multi-type power supply day-ahead demand forecasting system for scheduling and maintenance provided in an embodiment of the present invention; Figure 2 yes Figure 1 A schematic diagram of the unit of the intermediate coupling collaborative correction module; Figure 3 yes Figure 2 A schematic diagram of the sub-units of a typical scene generation unit. Detailed Implementation

[0047] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0048] like Figures 1 to 3 As shown in the figure, an embodiment of the present invention provides a multi-type power day-ahead demand forecasting system for scheduling and maintenance, the system comprising: The heterogeneous data processing module is used to collect various types of power source operation data, power load data, weather forecast data, and scheduling constraint data, and to preprocess them to generate time-series feature datasets.

[0049] The preprocessing steps performed by the heterogeneous data processing module include: The system collects operation sequence data of thermal power and hydropower, output data of wind power and photovoltaic power stations, operation data of energy storage charging and discharging, load data of the entire network and its zones, grid data of numerical weather forecasts, and maintenance data of generating units to form a raw multi-source dataset.

[0050] Multi-source data tables:

[0051] Outlier identification and removal are performed on the original multi-source dataset, missing data are repaired using nearest neighbor interpolation, and the scale of the entire dataset is normalized to obtain the basic dataset.

[0052] Outliers are identified using the 3σ criterion combined with box plots. For univariate time series... : Calculate the series mean and standard deviation :

[0053]

[0054] Judgment rule: If the data points satisfy If the value is not found, it is considered an outlier, marked, and removed.

[0055] For a single missing point, the formula is expressed as:

[0056] In the formula, It is a missing moment. , It refers to the valid time before and after. It is a missing moment The corresponding repaired data value, It is a missing moment Previous valid time The corresponding known data values, It is a missing moment Previous valid time The corresponding known data value.

[0057] For data missing for more than 3 consecutive time points, the historical 7-day average of the same type of unit / station during the same period is used to fill the gaps.

[0058] The basic dataset is timestamped using a preset standard time step to unify the data sampling frequency and obtain a time-aligned dataset.

[0059] Based on a scheduling standard of 15-minute steps, the basic dataset is resampled and aligned using timestamps to unify the sampling frequency.

[0060] Time base setting: Starting from 00:00 on a calendar day, according to... Divide the time points, the set of time points for a single day is as follows .

[0061] Downsampling: For data with a sampling frequency higher than 15 minutes (such as 1-minute unit data), the mean of the time period is aggregated.

[0062] Upsampling processing: For data with a sampling frequency of less than 15 minutes (such as hourly NWP data), cubic spline interpolation is used to complete the data to a 15-minute granularity. All data uses a unified UTC+8 time zone timestamp and is associated and matched according to the primary key (timestamp + site / unit ID) to ensure that multi-source data at the same point in time correspond one-to-one.

[0063] The steps of the heterogeneous data processing module in generating a time-series feature dataset include: Basic operational features are extracted from the time-series aligned dataset. Calendar features are constructed by combining date type, seasonal attributes, and holiday identifiers. Meteorological driving features are constructed by combining temperature, irradiance, wind speed, and cloud cover. Physical constraint features are constructed by combining unit rated parameters and grid structure parameters, forming a multi-dimensional feature factor library.

[0064] Feature factor table:

[0065] The multi-dimensional feature factor library is mapped one-to-one with the power output labels and load demand labels of the corresponding time period according to the timestamp, and a labeled feature sample set is constructed.

[0066] A rolling time window mechanism is used to divide the feature sample set into training, validation and test sets to generate a time series feature dataset.

[0067] Let the total sample length be N, the training window length be L, the prediction window length be H (H=96 in the current prediction), and the sliding step size be s: The k-th training set window: sample index is .

[0068] The k-th test set window: sample index is .

[0069] The validation set is split from the training set in an 8:2 ratio and used for hyperparameter tuning and early stopping.

[0070] The multi-type prediction module is used to read time-series feature datasets, classify them according to power supply operating characteristics, construct prediction units for each category, and output day-ahead prediction curves for each type of power supply in parallel.

[0071] The steps for the multi-type prediction module to output day-ahead prediction curves for various power sources include: The time-series feature dataset is divided into controllable power source subset, fluctuating new energy subset, energy storage operation subset, and power load subset according to power source controllability and output characteristics.

[0072] For a subset of controllable power sources, a time-series prediction model is constructed by combining the unit ramp rate and minimum output physical constraints, and the adjustable output baseline curves of thermal power and hydropower are output.

[0073] Set up the unit exist The output of time The constraints are as follows: Output upper and lower limit constraints:

[0074] in, To minimize the technical output of the unit, The rated active power output of the unit.

[0075] Climbing speed constraint:

[0076] in, , The unit's uphill and downhill ramp rates. It is the time step.

[0077] The total loss consists of two parts: the prediction error loss and the constraint penalty term, expressed by the formula:

[0078] in, It is the model predicting the output force. It is actual effort. yes The more limited the time constraint, It is the constraint penalty coefficient.

[0079] The output results are the baseline curves of adjustable output power for thermal and hydropower.

[0080] For a subset of fluctuating renewable energy sources, a decomposition-integration deep learning framework is adopted to construct an output prediction model by coupling numerical weather forecast features, and output day-ahead output prediction curves for wind power and photovoltaic power.

[0081] The steps for generating day-ahead power output forecast curves for wind and solar power include: The ensemble empirical mode decomposition algorithm is used to decompose the historical output sequence of new energy into low-frequency trend components that characterize long-term changes and high-frequency random components that characterize short-term fluctuations.

[0082] The historical power output sequence is analyzed using ensemble empirical mode decomposition (EEMD). Decomposed into several intrinsic mode functions (IMFs) and residual terms:

[0083] in, : No. The IMF components are arranged from highest to lowest frequency. It is the residual term, representing the long-term trend of output.

[0084] The components are reorganized into two categories based on a frequency threshold: Low-frequency trend component: includes residual terms and low-order IMF, representing seasonal and daily power output variation patterns.

[0085] High-frequency random components: contain higher-order IMFs, characterizing random fluctuations on the minute and hour levels.

[0086] The low-frequency trend component and the trend meteorological characteristics in numerical weather prediction are input into a temporal convolutional network to fit the long-term variation law of new energy output and generate low-frequency prediction components.

[0087] A Temporal Convolutional Network (TCN) is used to model the low-frequency trend components. The core is causal dilated convolution, and the output of the l-th convolutional layer is:

[0088] in, It is the coefficient of thermal expansion. It is the kernel size. , It refers to the kernel weights and biases. It is an activation function. The input is a low-frequency component plus trend meteorological features (daily average irradiance, daily average wind speed), and the output is a low-frequency prediction component.

[0089] The high-frequency random components and spatial correlation features of adjacent stations are input into a graph convolutional network to capture the wave transmission and spatial complementarity characteristics between stations in the region, and generate high-frequency prediction components.

[0090] Graph Convolutional Networks (GCNs) are used to capture spatial relationships between stations and construct a station adjacency matrix. (The elements are calculated based on the geographical distance between stations and their correlation with the terrain). The propagation formula for the convolutional layer is:

[0091] in, It is an adjacency matrix with self-loops added. yes The degree matrix, It is the first The node feature matrix of the layer It is the first The trainable weight matrix of the layer.

[0092] The input consists of high-frequency components and real-time meteorological data from adjacent stations, and the output consists of high-frequency forecast components.

[0093] By superimposing and reconstructing the low-frequency and high-frequency prediction components, the day-ahead power output prediction curves for wind power and photovoltaic power are obtained.

[0094] For a subset of energy storage operations, a boundary calculation model is constructed by combining initial state of charge, charge / discharge efficiency, and lifetime decay constraints, and the boundary curve of the chargeable / dischargeable power of the energy storage is output.

[0095] By combining the SOC evolution law and lifetime constraints, the maximum charge and discharge power boundary for each time period is calculated.

[0096] Energy storage State of charge at time t satisfy:

[0097] in, It is the self-discharge rate. , yes The charging and discharging power at any time It refers to charging and discharging efficiency. It is the rated capacity of energy storage. It refers to the efficiency of energy conversion during charging. It refers to the efficiency of energy conversion during discharge. yes The state of charge at any given moment.

[0098] Hard constraints SOC upper and lower limits: .

[0099] Maximum charging and discharging power: .

[0100] The equivalent cycle loss formula is calculated using the rainflow counting method to determine the lifetime loss corresponding to the cycle depth.

[0101] in, For the first Depth of discharge in the next cycle This is a depth-lifetime decay function. It is the number of loops. This refers to lifespan depletion. The total lifespan cost is amortized over a single cycle, serving as a soft constraint at the charge / discharge boundary.

[0102] The final output is the maximum charging power boundary and maximum discharging power boundary curves for each time period.

[0103] For a subset of electricity load, a load forecasting model is constructed by combining calendar features and meteorological sensitivity factors, and the day-ahead demand curve of regional electricity load is output.

[0104] A frequency decoupling modeling framework is adopted to decompose the load into basic load and meteorological sensitive load for separate prediction.

[0105] The formula for load decomposition is expressed as:

[0106] in, It is the base load component. It is a meteorologically sensitive load component. It is a random residual component.

[0107] Based on calendar features, a Fourier series is used to fit the periodic terms, expressed by the following formula:

[0108] in, It is the period length. These are Fourier coefficients. It is a date type feature vector. These are the corresponding coefficients. yes The basic load component at any given time. It is the harmonic sequence number.

[0109] The influence of meteorological factors such as temperature was fitted using multivariate nonlinear regression:

[0110] in, yes ambient temperature at all times It is the human body's comfort temperature benchmark. It is a positive part function. >0 time take Otherwise, take 0. It is the temperature sensitivity coefficient.

[0111] The final output is the day-ahead demand curve for regional electricity load.

[0112] The adjustable output baseline curve, day-ahead output forecast curve, chargeable and dischargeable power boundary curve, and regional power load day-ahead demand curve are spliced ​​and integrated according to a unified time step to form day-ahead forecast curves for various power sources.

[0113] The coupled collaborative correction module is used to mine spatiotemporal coupling correlation patterns based on the day-ahead prediction curves of various power sources, perform multivariate joint correction and uncertainty quantification, and generate a set of collaborative prediction scenarios with confidence intervals.

[0114] The coupling coordination correction module includes: The coupling correlation mining unit is used to read the day-ahead forecast curves of various power sources, extract the coupling features of wind-solar complementarity, source-load synchronization, and hydro-thermal substitution, and construct a multivariate correlation correction matrix.

[0115] Wind-solar complementarity: the degree of negative correlation between photovoltaic power output and wind power output.

[0116] Source-load synchronicity: The degree of positive correlation between photovoltaic output and air conditioning load (high temperature scenario).

[0117] Hydropower substitutability: The degree of negative correlation between hydropower output and thermal power output.

[0118] Build correlation correction matrix , To predict the total number of variables, matrix elements For variables With variables The normalized value of the coupling strength.

[0119] The collaborative baseline correction unit is used to perform spatiotemporal joint correction on the daily forecast curves through a multivariate correlation correction matrix, and generate a collaborative correction baseline curve.

[0120] Based on the correlation correction matrix, multivariate joint correction is performed on the independent prediction results to eliminate error accumulation. The correction formula is as follows:

[0121] in, It is an independent prediction result vector. It is the historical average prediction error vector of each variable. It is an error correction term based on the propagation of coupling relationships. It is a collaborative calibration reference curve.

[0122] The uncertainty quantification unit is used to statistically analyze the distribution characteristics of historical prediction errors, fit the error probability distribution function, and calculate the upper and lower bounds of the prediction interval at different confidence levels.

[0123] The probability density function of the error is fitted using nonparametric kernel density estimation (KDE). Given a confidence level (e.g., 95%), the upper and lower quantiles of the error distribution are calculated to obtain the upper and lower bounds of the prediction interval.

[0124] The typical scenario generation unit is used to generate an initial set of running scenarios based on the prediction interval, and after scenario reduction, typical running scenarios are extracted to generate a collaborative prediction scenario set.

[0125] Typical scene generation units include: The scene sampling unit is used to generate an initial set of running scenes based on the error probability distribution and prediction interval using the Latin hypercube sampling method. Each scene corresponds to a set of source load time series curves.

[0126] The cumulative probability interval [0,1] of the error distribution of each variable is divided into N equal probability sub-intervals. A sample is randomly drawn from each sub-interval. The sampling results of each variable are randomly combined to generate N initial running scenarios, which constitute the initial scenario set.

[0127] The scene reduction subunit is used to cluster and reduce the initial set of running scenarios using the K-means clustering algorithm, retaining high-probability normal scenarios and low-probability extreme risk scenarios to obtain typical running scenarios.

[0128] The scenario integration subunit is used to integrate and encapsulate typical operating scenarios with collaborative correction benchmark curves and upper and lower bounds of confidence intervals to generate a collaborative prediction scenario set.

[0129] The scheduling constraint adaptation module is used to embed the collaborative prediction scenario set into the hard constraints of power grid scheduling, perform feasibility verification and boundary correction, and generate multiple types of power demand prediction results.

[0130] The steps by which the scheduling constraint adaptation module generates multiple types of power demand forecasting results include: The baseline curve and time-period output and load values ​​of each typical scenario are extracted from the collaborative prediction scenario set.

[0131] The system incorporates unit operation constraints such as unit ramp rate, minimum technical output, and minimum start-up / shutdown time, as well as grid security constraints such as cross-sectional transmission limits and spinning reserve ratio, to construct a scheduling verification rule base.

[0132] Based on the scheduling verification rule base, the feasibility of the collaborative correction benchmark curve is verified on a time-by-time basis. For time periods exceeding the limit, the output boundary is smoothed and corrected to generate a benchmark prediction curve that meets the scheduling hard constraints.

[0133] Based on the 96-point time sequence, the collaborative correction baseline curve is verified to meet the rule base constraints for each time period, each power source, and each section: traversing each time point, the output and ramp-up constraints of a single unit are verified, the power balance and reserve constraints of the total output-total load of the region are verified, and the transmission power constraints of key sections are verified.

[0134] Minimum amplitude smoothing correction is applied to periods exceeding the limit, with the principle of preserving prediction accuracy as much as possible and satisfying all hard constraints.

[0135] Correction for exceeding climbing limits: Based on the output in the preceding and following periods, smoothly transition at the maximum climbing rate.

[0136] In accordance with the scheduling business specifications, the baseline forecast curve, confidence interval and typical scenarios are output in a preset time sequence format to generate multi-type power demand forecast results that can be directly used for day-ahead scheduling, categorized by region and power type.

[0137] According to the scheduling system data specifications, output the final result in the following format: Time granularity: 96 points per day, 15-minute step.

[0138] Data dimensions: by region and by power source type (thermal power, hydropower, wind power, photovoltaic, energy storage, load).

[0139] Data content: baseline predicted value, upper and lower bounds of 95% confidence interval, and curves for 5 typical scenarios.

[0140] Data format: Conforms to the general E-format / CIM model specification of the power system, and can be directly connected to the day-ahead unit combination (SCUC) and economic dispatch (SCED) models.

[0141] The rolling closed-loop iteration module is used to collect actual intraday operating data, calculate prediction deviations and identify the degree of data distribution drift, update the execution of each prediction unit online, and generate the iteratively optimized model parameter set.

[0142] The steps for the rolling closed-loop iterative module to identify the degree of data distribution drift include: The actual daily operation data is collected and compared point by point with the day-ahead forecast results for the corresponding time period to calculate the forecast deviation sequence by power source type and time period.

[0143] Based on the prediction bias sequence, the statistical error distribution characteristics of the sliding window are used, and the concept drift detection algorithm is used to determine the degree of deviation between the current data distribution and the training data distribution, and generate drift degree labels.

[0144] Set a sliding window of length W (usually 7 days), and statistically analyze the error distribution characteristics of the reference window (historical training data) and the current window respectively.

[0145] The Kolmogorov-Smirnov (KS) test was used to determine whether there was a significant difference in the error distributions of the two windows. Based on the magnitude of the KS statistic and the error increase, the degree of drift was divided into three levels: Slight drift: Error increase <5%, no immediate update required.

[0146] Moderate drift: Error increase of 5%~15%, triggering incremental update.

[0147] Severe drift: Error increase > 15%, triggering full retraining.

[0148] When the drift exceeds a preset threshold, an incremental update instruction is triggered to extract the latest time period's running data and construct an incremental training set.

[0149] The steps of the rolling closed-loop iteration module in generating the model parameter set include: The incremental training set is read, and the model parameters of each prediction unit are fine-tuned and optimized by using transfer learning and incremental training.

[0150] A transfer learning fine-tuning strategy is adopted to update the model parameters of each prediction unit based on the incremental training set: the parameters of the feature extraction network at the bottom layer of the model are frozen to retain the general feature extraction capability. Only the parameters of the top output layer and fully connected layer are updated to reduce the computational cost. The parameter update adopts the mini-batch gradient descent method.

[0151] The weight coefficients of each driving feature in the feature factor library are updated synchronously, and the feature importance ranking is corrected.

[0152] Based on incremental data, the importance of each driving feature is recalculated using SHAP values, expressed by the following formula:

[0153] in, For the first The SHAP value of each feature represents the contribution of that feature to the prediction result. Does not contain features a subset of It is the set of all features. It is a subset The corresponding model prediction value, Is it adding features? The predicted value after that.

[0154] The feature weights in the feature factor library are updated based on the calculation results, and the feature ranking is corrected. The optimized model parameters and feature weights are then solidified and encapsulated to generate a model parameter set, which is then distributed to each prediction unit to complete the model iteration.

[0155] In summary, this embodiment provides a multi-source power demand forecasting system for dispatching and operation, which significantly improves the accuracy and physical rationality of day-ahead forecasts for various power sources through heterogeneous data processing and differentiated modeling of different power types. It substantially improves the accuracy of day-ahead output forecasts for wind and solar power, especially under weather transitions and local fluctuations, where the predicted curves more closely resemble reality. Joint correction mitigates the cumulative effect of errors, and probability intervals and typical scenarios enable dispatching to make risk predictions and reserve decisions, enhancing the system's robustness against uncertainty. By embedding the forecast results into the hard constraints of grid dispatching for feasibility verification and correction, the generated power demand plan naturally meets the requirements for safe grid operation and can be directly imported into the dispatch automation system, greatly reducing manual intervention and improving the automation level and safety margin of dispatching and operation. Rolling closed-loop iteration and concept drift detection ensure the system's continued high accuracy and adaptability during long-term operation. It can quickly adapt to new operating modes without requiring full retraining, reducing operational complexity.

[0156] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these modifications or substitutions do 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 multi-type power supply day-ahead demand forecasting system for scheduling and operation and maintenance, characterized in that, include: The heterogeneous data processing module is used to collect various types of power source operation data, power load data, weather forecast data and scheduling constraint data, and perform preprocessing to generate time-series feature datasets. A multi-type prediction module is used to read the time-series feature dataset, classify it according to the power supply operating characteristics and construct prediction units for each category, and output the day-ahead prediction curves of each type of power supply in parallel. The coupled collaborative correction module is used to mine the spatiotemporal coupling correlation patterns based on the day-ahead prediction curves of various power sources, perform multivariate joint correction and uncertainty quantification, and generate a set of collaborative prediction scenarios with confidence intervals. The scheduling constraint adaptation module is used to embed the collaborative prediction scenario set into the grid scheduling hard constraints, perform feasibility verification and boundary correction, and generate multiple types of power demand prediction results. The rolling closed-loop iteration module is used to collect actual intraday operating data, calculate prediction deviations and identify the degree of data distribution drift, update the execution of each prediction unit online, and generate the iteratively optimized model parameter set.

2. The day-ahead demand forecasting system for multiple types of power sources oriented to scheduling and maintenance according to claim 1, characterized in that: The preprocessing steps performed by the heterogeneous data processing module include: The system collects operation sequence data of thermal power and hydropower, output data of wind power and photovoltaic power stations, operation data of energy storage charging and discharging, load data of the entire network and its zones, grid data of numerical weather forecasts, and maintenance data of generating units to form a raw multi-source dataset. Outlier identification and removal are performed on the original multi-source dataset, missing data are repaired using nearest neighbor interpolation, and the normalization of the dimensions of the entire dataset is completed to obtain the basic dataset. The basic dataset is timestamped using a preset standard time step to unify the data sampling frequency and obtain a time-aligned dataset.

3. A multi-type power supply day-ahead demand forecasting system for scheduling and maintenance according to claim 2, characterized in that: The steps of the heterogeneous data processing module in generating the time-series feature dataset include: Basic operational features are extracted from the time-series aligned dataset. Calendar features are constructed by combining date type, seasonal attributes, and holiday identifiers. Meteorological driving features are constructed by combining temperature, irradiance, wind speed, and cloud cover. Physical constraint features are constructed by combining unit rated parameters and grid structure parameters, forming a multi-dimensional feature factor library. The multi-dimensional feature factor library is mapped and associated one-to-one with the power output label and load demand label of the corresponding time period according to the timestamp, and a labeled feature sample set is constructed. The feature sample set is divided into training set, validation set and test set using a rolling time window mechanism to generate the time-series feature dataset.

4. A multi-type power supply day-ahead demand forecasting system for scheduling and maintenance according to claim 1, characterized in that: The steps for the multi-type prediction module to output day-ahead prediction curves for various power sources include: The time-series feature dataset is divided into a controllable power source subset, a fluctuating new energy source subset, an energy storage operation subset, and an electricity load subset based on power source controllability and output characteristics. For the controllable power source subset, a time-series prediction model is constructed by combining the unit ramp rate and minimum output physical constraints, and the adjustable output baseline curves of thermal power and hydropower are output. For the aforementioned subset of volatile new energy sources, a decomposition-integration deep learning framework is adopted, coupled with numerical weather forecast features to construct an output prediction model, and output day-ahead output prediction curves for wind power and photovoltaic power. For the aforementioned energy storage operation subset, a boundary calculation model is constructed by combining the initial state of charge, charge / discharge efficiency, and lifetime decay constraints, and the rechargeable / dischargeable power boundary curve of the energy storage is output. For the aforementioned subset of electricity load, a load forecasting model is constructed by combining calendar features and meteorological sensitivity factors, and the day-ahead demand curves for regional electricity loads are output. The adjustable output baseline curve, day-ahead output forecast curve, chargeable and dischargeable power boundary curve, and regional power load day-ahead demand curve are spliced ​​and integrated according to a unified time step to form day-ahead forecast curves for various power sources.

5. A multi-type power day-ahead demand forecasting system for scheduling and maintenance according to claim 4, characterized in that: The steps for generating day-ahead power output forecast curves for wind and solar power include: The ensemble empirical mode decomposition algorithm is used to decompose the historical output sequence of new energy into low-frequency trend components that characterize long-term changes and high-frequency random components that characterize short-term fluctuations. The low-frequency trend component and the trend meteorological characteristics in numerical weather forecasting are input into a temporal convolutional network to fit the long-term variation law of new energy output and generate low-frequency prediction components. The high-frequency random component and the spatial correlation features of adjacent stations are input into a graph convolutional network to capture the wave transmission and spatial complementarity characteristics between stations in the region, and generate high-frequency prediction components. The low-frequency prediction component and the high-frequency prediction component are superimposed and reconstructed to obtain the day-ahead power output prediction curves of wind power and photovoltaic power.

6. A multi-type power day-ahead demand forecasting system for scheduling and maintenance according to claim 1, characterized in that: The coupling coordination correction module includes: The coupling correlation mining unit is used to read the day-ahead forecast curves of various power sources, extract the coupling features of wind-solar complementarity, source-load synchronization, and hydro-thermal substitution, and construct a multivariate correlation correction matrix. The collaborative benchmark correction unit is used to perform spatiotemporal joint correction on the daily forecast curves through the multivariate correlation correction matrix to generate a collaborative correction benchmark curve. Uncertainty quantification unit is used to statistically analyze the distribution characteristics of historical prediction errors, fit the error probability distribution function, and calculate the upper and lower bounds of the prediction interval at different confidence levels. The typical scenario generation unit is used to generate an initial set of running scenarios based on the prediction interval, and after scenario reduction, extract typical running scenarios to generate the collaborative prediction scenario set.

7. A multi-type power supply day-ahead demand forecasting system for scheduling and maintenance according to claim 6, characterized in that: The typical scene generation unit includes: The scene sampling unit is used to generate the initial running scene set based on the error probability distribution and prediction interval using the Latin hypercube sampling method. Each scene corresponds to a set of source load time series curves. The scenario reduction subunit is used to cluster and reduce the initial set of running scenarios using the K-means clustering algorithm, retaining high-probability regular scenarios and low-probability extreme risk scenarios to obtain the typical running scenarios. The scenario integration subunit is used to integrate and encapsulate the typical operating scenario with the collaborative correction benchmark curve and the upper and lower bounds of the confidence interval to generate the collaborative prediction scenario set.

8. A multi-type power day-ahead demand forecasting system for scheduling and maintenance according to claim 7, characterized in that: The steps of the scheduling constraint adaptation module in generating multiple types of power demand forecasting results include: The baseline curve and the time-by-time output and load values ​​of each typical scenario are extracted from the collaborative prediction scenario set; The system incorporates unit operation constraints such as unit ramp rate, minimum technical output, and minimum start-up and shutdown time, as well as grid security constraints such as cross-sectional transmission limit and spinning reserve ratio, to construct a scheduling verification rule base. Based on the scheduling verification rule base, the feasibility of the collaborative correction benchmark curve is verified time by time. For time periods exceeding the limit, the output boundary is smoothed and corrected to generate a benchmark prediction curve that meets the scheduling hard constraints. In accordance with the scheduling business specifications, the baseline forecast curve, confidence interval and typical scenarios are output in a preset time sequence format to generate multi-type power demand forecast results that can be directly used for day-ahead scheduling, categorized by region and power type.

9. A multi-type power day-ahead demand forecasting system for scheduling and maintenance according to claim 3, characterized in that: The steps for the rolling closed-loop iteration module to identify the degree of data distribution drift include: The actual daily operation data is collected and compared point by point with the day-ahead forecast results for the corresponding time period to calculate the forecast deviation sequence by power source type and time period. Based on the predicted deviation sequence, the sliding window statistical error distribution characteristics are used, and the concept drift detection algorithm is used to determine the degree of deviation between the current data distribution and the training data distribution, and a drift degree label is generated. When the drift exceeds a preset threshold, an incremental update instruction is triggered to extract the latest time period's running data and construct an incremental training set.

10. A multi-type power day-ahead demand forecasting system for scheduling and maintenance according to claim 9, characterized in that: The steps of generating the model parameter set by the rolling closed-loop iteration module include: The incremental training set is read, and the model parameters of each prediction unit are fine-tuned and optimized using transfer learning and incremental training methods. The weight coefficients of each driving feature in the feature factor library are updated synchronously to correct the feature importance ranking; The optimized model parameters and feature weights are solidified and encapsulated to generate the model parameter set, which is then distributed to each prediction unit to complete the model iteration.

Images