CN-121983964-A - Photovoltaic power generation ultra-short-term power prediction method and system based on LSTM-converter mixed model

CN121983964ACN 121983964 ACN121983964 ACN 121983964ACN-121983964-A

Abstract

The application provides a photovoltaic power generation ultra-short-term power prediction method and system based on an LSTM-converter mixed model, and relates to the field of power system operation control. The method comprises the steps of constructing an evolution rate time sequence based on meteorological time sequence data, constructing a time sequence evolution-aware LSTM model by taking the evolution rate time sequence as a gating adjustment factor, extracting the time sequence characteristics of an evolution state, inputting the time sequence characteristics of the evolution state, a stable trend component decomposed based on historical power time sequence data, a short-time disturbance component and a disturbance intensity weight sequence into a transducer model, introducing the evolution rate time sequence into a self-attention mechanism of the transducer model as a constraint factor, extracting the correlation characteristics of photovoltaic power, and predicting the photovoltaic power generation power. The method is used for the ultra-short term power prediction process of the photovoltaic power generation, and solves the technical problem of error accumulation caused by prediction lag in the existing ultra-short term photovoltaic power generation power prediction process.

Inventors

WANG YI
YANG DONG
GAO JIRONG
LI ZHENRONG
YANG JIANNENG
BAI YAN
LIU ZHEN

Assignees

华能澜沧江新能源有限公司

Dates

Publication Date: 20260505
Application Date: 20260204

Claims (10)

1. A photovoltaic power generation ultra-short-term power prediction method based on an LSTM-converter mixed model is characterized by comprising the following steps of: The method comprises the steps of obtaining multi-source time sequence data of a photovoltaic power station, wherein the multi-source time sequence data of the photovoltaic power station comprises historical power time sequence data and meteorological time sequence data of the photovoltaic power station; calculating the variation amplitude of the meteorological variable at adjacent moments based on the meteorological time sequence data, and constructing an evolution rate time sequence; taking the multi-source time sequence data of the photovoltaic power station as input, constructing a time sequence evolution perceived LSTM model by taking an evolution rate sequence as a gating adjustment factor, and extracting the time sequence characteristics of an evolution state; Decomposing the historical power time sequence data into a stable trend component and a short-time disturbance component by a time sequence difference method, and calculating a disturbance intensity weight sequence; Inputting the evolution state time sequence characteristics, the stable trend component, the short-time disturbance component and the disturbance intensity weight sequence into a transducer model, introducing an evolution rate time sequence into a self-attention mechanism of the transducer model as a constraint factor, and extracting the photovoltaic power correlation characteristics; and based on the photovoltaic power correlation characteristics, predicting the photovoltaic power generation power within a preset time step through a prediction layer.
2. The method of claim 1, wherein calculating the magnitude of change of the weather variable at adjacent times based on the weather timing data, and constructing the evolution rate timing sequence, comprises: Acquiring weather time sequence data of adjacent moments, and carrying out standardized processing on weather variables, wherein the weather variables are solar irradiance, cloud cover and environmental temperature in the weather time sequence data; Calculating the change amplitude of the meteorological variable based on the difference value of the meteorological variable at the adjacent moment, and carrying out weighted fusion on the change amplitude of the meteorological variable according to a preset weight to obtain the comprehensive change quantity; And carrying out normalization and/or smoothing treatment on the comprehensive variation to obtain a time sequence of the time sequence evolution rate.
3. The method of claim 1, wherein the constructing a time-series evolution-aware LSTM model with the photovoltaic power plant multisource time-series data as input and the evolution rate sequence as a gating regulator, extracting the evolution state time-series sequence features, comprises: Inputting the multi-source time sequence data of the photovoltaic power station into an LSTM network perceived by time sequence evolution according to time sequence to obtain a time sequence hidden state; Introducing an evolution rate sequence into a gating structure of an LSTM network perceived by time sequence evolution, regulating forgetting gate, input gate and state updating process, and outputting the time sequence characteristics of the evolution state.
4. The method of claim 3, wherein the adjusting gating of the LSTM model of time-sequence evolution perception comprises an evolution perception forget gate, an evolution perception input gate and an evolution perception output gate; the evolution perception forgetting door The following formula is satisfied: Wherein, the For the LSTM hidden state perceived by the time sequence evolution at the last moment, In order to evolve the perceived forgetting gate weight matrix, In order to evolve the perceived forgetting gate bias term, Is the input vector at the time t, In order to adjust the function of the evolution rate, For the sequence of evolution rates, Activating a function for sigmoid, wherein t represents a time step index; the evolution-aware input gate satisfies the following formula: Wherein, the For the LSTM hidden state perceived by the time sequence evolution at the last moment, To evolve the perception input gate weight matrix, To evolve the sense input gate bias term, Is the input vector at the time t, In order to adjust the function of the evolution rate, For the sequence of evolution rates, Activating a function for sigmoid, wherein t represents a time step index; The evolution-aware output gate satisfies the following formula: Wherein, the For the LSTM hidden state perceived by the time sequence evolution at the last moment, The gate weight matrix is output for evolutionary perception, The gate bias term is output for evolution awareness, Is the input vector at the time t, In order to adjust the function of the evolution rate, For the sequence of evolution rates, For sigmoid activation functions, t represents a time step index.
5. The method of claim 1, wherein the decomposing the historical power timing data into a steady trend component and a short disturbance component by a timing difference method and calculating a disturbance intensity weight sequence comprises: Calculating the power difference value of the adjacent moments of the historical power time sequence data according to the time sequence; Decomposing the historical power time sequence data into a stable trend component and a short-time disturbance component based on the power difference value; And calculating a disturbance intensity weight sequence according to the short-time disturbance component.
6. The method of claim 5, wherein the steady trend component The following formula is satisfied: Wherein, the For photovoltaic power at time t+i, k is the half width of the sliding window, and t represents the time step index.
7. The method of claim 1, wherein said inputting the evolution state timing sequence feature, the steady trend component, the short term disturbance component, and the disturbance intensity weight sequence into a transducer model and introducing an evolution rate timing sequence as a constraint factor in a self-attention mechanism of the transducer model, extracting the photovoltaic power correlation feature comprises: Performing sequence splicing on the evolution state time sequence features, the stable trend component, the short-time disturbance component and the disturbance intensity weight sequence according to a time step to form a multichannel input sequence; mapping the multichannel input sequence into a query sequence, a key sequence and a value sequence respectively, and inputting the self-attention layer of the transducer model; Generating a constraint matrix based on the evolution rate time sequence, and adding the constraint matrix into an attention calculation process between the query sequence and the key sequence to obtain attention weights; And carrying out weighted fusion on the multichannel input sequences according to the attention weight to obtain a photovoltaic power association characteristic sequence.
8. The method of claim 7, wherein the photovoltaic power-related signature sequence The following formula is satisfied: Wherein, the For the purpose of the query sequence, A sequence of the keys is provided, The sequence of values is used to determine, In order to evolve the rate constraint matrix, In order for the scaling factor to be a factor, Is a weight coefficient.
9. The method of claim 1, wherein predicting, by a prediction layer, photovoltaic power generation for a preset time step based on the photovoltaic power correlation characteristic, comprises: inputting the photovoltaic power association characteristic sequence into a full-connection prediction layer to obtain a preliminary power prediction value corresponding to each time step; and inputting the preliminary power predicted value into an activation function layer to obtain the photovoltaic power generation power in a preset time step.
10. The photovoltaic power generation ultra-short-term power prediction system based on the LSTM-converter mixed model is characterized by comprising a data acquisition module, an evolution rate construction module, a time sequence evolution perception LSTM feature extraction module, a power time sequence decomposition module, an evolution rate constraint converter feature fusion module and a power prediction module; The data acquisition module is used for acquiring multi-source time sequence data of the photovoltaic power station, wherein the multi-source time sequence data comprises historical power time sequence data and meteorological time sequence data of the photovoltaic power station; The evolution rate construction module is used for calculating the variation amplitude of the meteorological variable at adjacent moments based on the meteorological time sequence data and constructing an evolution rate time sequence, wherein the meteorological variable comprises solar irradiance, cloud cover and environmental temperature; The time sequence evolution perception LSTM characteristic extraction module is used for taking the multi-source time sequence data of the photovoltaic power station as input, taking the evolution rate time sequence as a gating adjustment factor, constructing a time sequence evolution perception LSTM model and outputting the time sequence characteristics of the evolution state; The power time sequence decomposition module is used for processing the historical power time sequence data through a time sequence difference method, decomposing to obtain a stable trend component and a short-time disturbance component, and calculating a disturbance intensity weight sequence; The evolution rate constraint converter feature fusion module is used for carrying out sequence splicing on the evolution state time sequence features, the stable trend components, the short-time disturbance components and the disturbance intensity weight sequences according to time steps to form a multichannel input sequence, generating a constraint matrix according to the evolution rate time sequence, introducing the constraint matrix into a self-attention computing process, and outputting a photovoltaic power association feature sequence; The power prediction module is used for outputting a photovoltaic power generation power prediction result within a preset time step through a prediction layer based on the photovoltaic power association characteristic sequence.

Description

Photovoltaic power generation ultra-short-term power prediction method and system based on LSTM-converter mixed model Technical Field The application relates to the field of operation control of power systems, in particular to a photovoltaic power generation ultra-short-term power prediction method and system based on an LSTM-converter mixed model. Background Photovoltaic power generation power prediction is an important technical foundation for guaranteeing safety, stability and economic operation of a power grid. Because the photovoltaic power generation power has high sensitivity to external environment changes in an ultra-short time scale, the power time sequence data of the photovoltaic power generation power has the characteristics of frequent fluctuation, obvious nonlinear characteristics, obvious local abrupt change and the like, so that modeling difficulty of ultra-short-term power prediction is increased. The existing photovoltaic power generation power prediction method mainly uses historical power time sequence data as a main modeling basis, and focuses on describing the overall evolution rule of power change along with time. In the ultra-short-term prediction application scene, when the power sequence is influenced by a rapid change factor, the method is difficult to reflect the instantaneous change characteristic of the power in time, and the phenomenon of lagging of a prediction result is generated, so that the problem of continuous accumulation of prediction errors is caused. Therefore, how to effectively solve the problem of error accumulation caused by prediction lag in the ultra-short-term photovoltaic power generation power prediction process has become a technical problem to be solved in the prior art. Disclosure of Invention The application provides a photovoltaic power generation ultra-short-term power prediction method and system based on an LSTM-converter mixed model, which solve the technical problem of error accumulation caused by prediction lag in the existing ultra-short-term photovoltaic power generation power prediction process. In order to achieve the above purpose, the application adopts the following technical scheme: According to the method, the system comprises the steps of obtaining multi-source time sequence data of a photovoltaic power station, wherein the multi-source time sequence data of the photovoltaic power station comprise historical power time sequence data of the photovoltaic power station and weather time sequence data, calculating change amplitude of weather variables at adjacent moments based on the weather time sequence data, constructing an evolution rate time sequence, taking the multi-source time sequence data of the photovoltaic power station as input, constructing a time sequence evolution perceived LSTM model by taking the evolution rate sequence as a gating regulator, extracting evolution state time sequence characteristics, decomposing the historical power time sequence data into stable trend components and short-time disturbance components through a time sequence difference method, calculating disturbance intensity weight sequences, inputting the evolution state time sequence characteristics, the stable trend components, the short-time disturbance components and the disturbance intensity weight sequences into the Transformer model, introducing the evolution rate time sequence into an independent attention mechanism of the Transformer model as a constraint factor, extracting photovoltaic power correlation characteristics, and predicting photovoltaic power in a preset time step length through a prediction layer based on the photovoltaic power correlation characteristics. According to the technical scheme, in the ultra-short-term power prediction method for photovoltaic power generation based on the LSTM-transducer mixed model, by introducing an evolution rate time sequence constructed by meteorological time sequence data in a multi-source time sequence data modeling process, the model can sense dynamic change characteristics of photovoltaic power generation under an ultra-short-term time scale and adaptively model a power evolution process according to the dynamic change characteristics, meanwhile, the LSTM model perceived by time sequence evolution is combined to extract the power time sequence characteristics, trend and disturbance component decomposition is carried out on historical power data, the influence of short-term fluctuation and random disturbance on a prediction result is effectively reduced, further, correlation modeling among multiple types of time sequence characteristics is realized by introducing evolution rate constraint in a transducer self-attention mechanism, the characteristic expression capacity of key time steps is enhanced, so that the problems of common prediction lag and error accumulation in ultra-short-term power prediction are effectively relieved, and the accuracy and stability of ultra-short-term power prediction of