A method and system for short-term net load forecasting that takes into account seasonal characteristics

Abstract

This invention discloses a short-term net load forecasting method and system that takes into account seasonal characteristics, belonging to the field of power load forecasting. The method includes: extracting features from historical photovoltaic (PV) output data and historical meteorological data to obtain an initial feature time series sequence representing the correlation between meteorological factors and PV output; inputting the initial feature time series sequence and seasonal feature codes into a BiLSTM to obtain an enhanced feature time series sequence representing the correlation between meteorological factors and PV output with integrated seasonal characteristics; processing the enhanced feature time series sequence through an attention mechanism to obtain the final weights of meteorological factors for PV output under different seasons; constructing a meteorological forecasting model; using the meteorological forecasting model to obtain predicted meteorological data; obtaining predicted PV output data based on the predicted meteorological data and the final weights; and obtaining the predicted net load based on the predicted PV output data. This invention solves the technical problem of existing technologies failing to improve the accuracy of PV output forecasting, leading to low accuracy in net load forecasting.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting, specifically to a short-term net load forecasting method and system that takes into account seasonal characteristics. Background Technology

[0002] The power system is undergoing profound changes, accelerating its transition towards higher proportions of renewable energy and greater power electronics. Short-term net load forecasting, as a core technology supporting grid dispatch and control optimization, directly impacts the balance of power supply and demand, economic dispatch, and safe and stable operation. Photovoltaic (PV) power generation, as a clean and sustainable energy form, relies heavily on output stability for accurate short-term net load forecasting. However, the stability of PV output is closely linked to meteorological conditions, making meteorological factors key variables affecting the efficient utilization of PV energy. Sunlight intensity and sunshine duration, as the direct energy source for PV power generation, directly determine the amount of energy received by the PV panels, thus significantly affecting the power generation of the PV system. To more accurately predict PV output data, existing technologies consider the impact of meteorological factors on PV output, calculating the influence weights of each meteorological factor and combining this with meteorological data for forecasting. However, existing technologies often assume that the impact of meteorological factors on PV output remains constant throughout the year, which clearly deviates from reality. In fact, the impact of meteorological factors on PV output is profoundly constrained by seasonal factors, with significant differences in the influence of various meteorological factors across different seasons. For example, in summer, the solar altitude angle is high and the light intensity is high, but the high temperature may cause a significant decrease in the efficiency of photovoltaic modules. At this time, the impact of temperature on photovoltaic output is more prominent. In winter, the sunshine duration is shorter and the light intensity is relatively weaker, making the constraint of light on photovoltaic output more critical. Therefore, how to construct the potential correlation between different meteorological factors and photovoltaic output in different seasons to improve the accuracy of photovoltaic output forecasting, and thus improve the accuracy of net load forecasting, is a technical challenge that urgently needs to be solved by existing technologies. Summary of the Invention

[0003] To address the technical problem of low net load forecast accuracy due to the difficulty in improving the accuracy of photovoltaic (PV) output forecasting in existing technologies, this invention provides a short-term net load forecasting method and system that takes into account seasonal characteristics. The method involves inputting an initial feature time series characterizing the correlation between meteorological factors and PV output, along with feature codes characterizing the season, into a BiLSTM to obtain an enhanced feature time series characterizing the potential correlation between different meteorological factors and PV output in different seasons. An attention mechanism is then used to process the enhanced feature time series to obtain the final weights of meteorological factors for PV output in different seasons. Based on the predicted meteorological data and the final weights, PV output forecast data is obtained. This solution addresses the technical problem of low net load forecast accuracy due to the difficulty in improving the accuracy of PV output forecasting in existing technologies.

[0004] To address the aforementioned technical problems, this invention provides a short-term net load forecasting method that takes into account seasonal characteristics, comprising the following steps:

[0005] Feature extraction was performed on historical photovoltaic power output data and historical meteorological data to obtain an initial feature time series that characterizes the correlation between meteorological factors and photovoltaic power output;

[0006] The initial feature time series sequence and the feature encoding representing the season are input into BiLSTM to obtain the enhanced feature time series sequence representing the correlation between meteorological factors and photovoltaic output, which represents the fused seasonal characteristics;

[0007] The attention mechanism is used to process the time series of enhanced features to obtain the final weights of meteorological factors on photovoltaic output under different seasons;

[0008] A meteorological forecasting model is constructed, and the forecasted meteorological data is obtained using the meteorological forecasting model. Based on the forecasted meteorological data and the final weights, the forecasted photovoltaic power output data is obtained, and the forecasted net load is obtained based on the forecasted photovoltaic power output data.

[0009] Preferably, the step of extracting features from historical photovoltaic power output data and historical meteorological data to obtain an initial feature time series characterizing the correlation between meteorological factors and photovoltaic power output includes:

[0010] Time alignment and standardization processes are performed on historical photovoltaic power output data and historical meteorological data to obtain a suitable network structure;

[0011] Initial correlation features are obtained by sliding convolutional layers in a CNN along a preset time step on the adapted network structure to capture the local correlation between changes in meteorological factors and changes in photovoltaic power output.

[0012] The initial associated features are nonlinearly transformed by the activation function in the CNN to obtain nonlinear associated features. Key associated features in the nonlinear associated features are then selected along a preset time step to obtain the time sequence of the initial features.

[0013] Preferably, the step of inputting the initial feature time series sequence and the feature encoding representing the season into a BiLSTM to obtain the enhanced feature time series sequence representing the correlation between meteorological factors and photovoltaic output, which characterizes the fused seasonal characteristics, includes:

[0014] The initial feature time series is divided into a first initial feature time series and a second initial feature time series. Based on the seasonal characteristics and the first initial feature time series, the conventional single-path forget gate in BiLSTM is transformed into a seasonal enhanced forget gate. The second initial feature time series and the feature encoding are input into BiLSTM to obtain the enhanced feature time series.

[0015] Preferably, the modification of the conventional single-path forget gate in BiLSTM into a seasonally enhanced forget gate based on seasonal characteristics and the time-series sequence of the first initial features includes:

[0016] Based on the fluctuation trend of historical meteorological data in different seasons, meteorological thresholds for historical meteorological data in different seasons are obtained, and seasonal calibration items are obtained based on the meteorological thresholds.

[0017] A seasonal adaptation model is constructed based on seasonal calibration items and preset initial seasonal weights. The spliced ​​first initial feature time sequence and feature encoding are used as inputs to train the seasonal adaptation model to obtain the process seasonal weights.

[0018] If the process seasonal weights meet the preset conditions, the seasonal adaptation model is updated using the process seasonal weights. Based on the seasonal adaptation model, the conventional single-path forget gate is transformed into a seasonally enhanced forget gate. Otherwise, the process seasonal weights are used as the preset initial seasonal weights to obtain the seasonally enhanced forget gate.

[0019] Preferably, the step of processing the enhanced feature time series through an attention mechanism to obtain the final weights of meteorological factors for photovoltaic output in different seasons includes:

[0020] The seasonal coding vector in the enhanced feature time series is mapped to the query vector, the meteorological factor feature vector in the enhanced feature time series is mapped to the key vector, and the correlation feature vector between meteorological factors and photovoltaic output in the enhanced feature time series is mapped to the value vector. The similarity between the query vector and the key vector is calculated, and the final weight of meteorological factors for photovoltaic output in different seasons is obtained based on the similarity and the value vector.

[0021] Preferably, the step of constructing a meteorological forecasting model and using the meteorological forecasting model to obtain forecast meteorological data includes:

[0022] A weather forecasting model is constructed based on CNN and BiLSTM. Historical weather data is input into the weather forecasting model to obtain estimated weather data. The weather forecasting model is then corrected based on the difference between the estimated weather data and the actual weather data. Predicted weather data is then obtained based on the corrected weather forecasting model.

[0023] Preferably, the step of obtaining photovoltaic power output forecast data based on predicted meteorological data and final weights includes:

[0024] The weight of the predicted meteorological data for photovoltaic power output is obtained based on the final weight, and the predicted photovoltaic power output data is obtained based on the weight of the predicted meteorological data for photovoltaic power output and the predicted meteorological data.

[0025] By adopting the above technical solution, the present invention has the following advantages:

[0026] Considering that the conventional single-channel forget gate in BiLSTM lacks a seasonal constraint mechanism, its feature selection process lacks adaptability to seasonal differences, making it difficult to uncover the potential correlation between different meteorological factors and photovoltaic power generation in different seasons. This makes it difficult to capture the dynamic changes in the weights of meteorological factors caused by seasonal transitions. Therefore, by using seasonal characteristics and the first initial feature time series, the conventional single-channel forget gate in BiLSTM is transformed into a seasonally enhanced forget gate. The second initial feature time series and the feature encoding representing the season are input into the seasonally enhanced forget gate to obtain an enhanced feature time series representing the correlation between meteorological factors with integrated seasonal characteristics and photovoltaic output. Based on this, an attention mechanism is introduced to perform refined weighting processing on this enhanced sequence, adaptively allocating the final weights of each meteorological factor on photovoltaic output in different seasons. Finally, the predicted photovoltaic output data is obtained by combining the meteorological data of the prediction period with the final weight matrix, and the predicted net load is obtained based on the predicted photovoltaic output data. This solves the technical problem that existing technologies struggle to improve the accuracy of photovoltaic output prediction, leading to low accuracy in net load prediction.

[0027] By constructing a weather forecasting model using CNN and BiLSTM, the model retains the ability of CNN to capture local weather changes while leveraging BiLSTM to capture the long-term temporal dependencies of weather data. Combined with an estimation-correction mechanism, this makes the predicted weather data more closely reflect reality. In calculating photovoltaic power output forecasts, instead of using a uniform year-round weather weight, the model directly calls the seasonally specific weather weights corresponding to the forecast period. This avoids the adaptation bias caused by the traditional one-size-fits-all approach to weights, ultimately effectively improving the accuracy of short-term net load forecasts.

[0028] This invention also provides a short-term net load forecasting system that takes into account seasonal characteristics, applicable to the aforementioned short-term net load forecasting method that takes into account seasonal characteristics, comprising:

[0029] The initial feature time series acquisition module is used to extract features from historical photovoltaic power output data and historical meteorological data to obtain initial feature time series that characterize the correlation between meteorological factors and photovoltaic power output;

[0030] An enhanced feature time series acquisition module is used to input the initial feature time series and the feature encoding representing the season into BiLSTM to obtain an enhanced feature time series that represents the correlation between meteorological factors and photovoltaic output, which represents the fused seasonal characteristics.

[0031] The meteorological factor weight acquisition module is used to process the enhanced feature time series through an attention mechanism to obtain the final weight of meteorological factors for photovoltaic power output under different seasons.

[0032] The predicted net load acquisition module is used to construct a meteorological forecasting model, use the meteorological forecasting model to acquire predicted meteorological data, acquire photovoltaic predicted output data based on the predicted meteorological data and the final weights, and acquire the predicted net load based on the photovoltaic predicted output data.

[0033] By adopting the above technical solution, the present invention has the following advantages:

[0034] By inputting the initial feature time series representing the correlation between meteorological factors and photovoltaic power output and the feature encoding representing the season into BiLSTM, an enhanced feature time series representing the potential correlation between different meteorological factors and photovoltaic power output in different seasons is obtained. The enhanced feature time series is then processed through an attention mechanism to obtain the final weights of meteorological factors for photovoltaic power output in different seasons. Based on the predicted meteorological data and the final weights, the predicted photovoltaic power output data is obtained. This solves the technical problem that existing technologies are unable to improve the accuracy of photovoltaic power output prediction, resulting in low accuracy of net load prediction.

[0035] The present invention also provides a computer device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the short-term net load forecasting method taking into account seasonal characteristics.

[0036] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the aforementioned short-term net load forecasting method taking into account seasonal characteristics. Attached Figure Description

[0037] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0038] Figure 1 This is a flowchart illustrating a short-term net load forecasting method that takes into account seasonal characteristics, according to the present invention.

[0039] Figure 2 This is a comparison chart of the first effects of a short-term net load forecasting method that takes into account seasonal characteristics according to the present invention;

[0040] Figure 3 This is a comparison diagram of the second effect of a short-term net load forecasting method that takes into account seasonal characteristics according to the present invention;

[0041] Figure 4This is a comparison chart of the third effect of the short-term net load forecasting method that takes into account seasonal characteristics according to the present invention.

[0042] Figure 5 This is a fourth comparison chart of the effects of the short-term net load forecasting method that takes into account seasonal characteristics of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0044] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0045] Example 1:

[0046] like Figure 1 As shown, a short-term net load forecasting method that takes into account seasonal characteristics includes the following steps:

[0047] S1: Extract features from historical photovoltaic power output data and historical meteorological data to obtain an initial feature time series that characterizes the relationship between meteorological factors and photovoltaic power output.

[0048] The process of extracting features from historical photovoltaic power output data and historical meteorological data to obtain an initial feature time series sequence characterizing the correlation between meteorological factors and photovoltaic power output includes:

[0049] Time alignment and standardization processes are performed on historical photovoltaic power output data and historical meteorological data to obtain a suitable network structure;

[0050] Initial correlation features are obtained by sliding convolutional layers in a CNN along a preset time step on the adapted network structure to capture the local correlation between changes in meteorological factors and changes in photovoltaic power output.

[0051] The initial associated features are nonlinearly transformed by the activation function in the CNN to obtain nonlinear associated features. Key associated features in the nonlinear associated features are then selected along a preset time step to obtain the time sequence of the initial features.

[0052] In this embodiment, historical meteorological data specifically refers to temperature, radiation, wind speed, precipitation, etc. Time alignment and standardization processing are performed on historical photovoltaic output data and historical meteorological data to obtain an adapted network structure. This includes: obtaining a time alignment benchmark based on net load forecast demand; using this benchmark, performing time alignment processing on the historical photovoltaic output data and historical meteorological data to obtain aligned data; and standardizing the aligned data based on the number of aligned data, the dimensions of the historical meteorological data, and the dimensions of the historical photovoltaic output data to obtain an adapted network structure. For example, if temperature, radiation, wind speed, and precipitation are used as historical meteorological data, and the "15-minute time stamp" is used as the time alignment benchmark, matching the photovoltaic output data and meteorological data one-to-one by timestamp, and deleting output records without corresponding meteorological data, a final result of 69,952 aligned data entries with completely identical timestamps is obtained. The adapted network structure is then a 69,952×5 data matrix. Local correlations can specifically be categorized as "continuous increase in radiation intensity → synchronous increase in photovoltaic output," "a slight increase in temperature has a positive auxiliary correlation on output," and "wind speed has a weak impact on output when there is no precipitation." Each convolutional kernel in the convolutional layer outputs a local correlation. If the convolutional layer contains 32 kernels, the initial correlation features are a 69952×32 matrix. It is understandable that different substations will have different local correlations due to differences in geographical location (e.g., the presence of surrounding buildings for obstruction) and equipment characteristics (e.g., photovoltaic array tilt angle, bracket installation method, and whether a cooling system is provided). The mathematical expression for the convolution operation of the convolutional layer is: , This represents the element at position (i, j) in the initial associated features after the convolution operation. This represents the element at position (i+m, j+n) in the adaptation network structure, where i and j are the position indices of the adaptation network structure, and m and n are the position indices of the convolutional kernels. This represents the weight parameters in the convolution kernel located at position (m, n). The bias term of the convolutional layer is used to adjust the offset of the result of the convolution operation. M and N represent the size of the convolution kernel in two dimensions, respectively, and together they define the size of the convolution kernel. The activation function can be the Leaky ReLU activation function, which performs a nonlinear transformation on the initial associated features, filtering out meaningless weakly associated negative feature values ​​while retaining some weak positive associations. The preset time step can be 1, i.e., moving one convolution kernel at a time, or it can be flexibly set according to user needs. In the pooling layer of the CNN (Convolutional Neural Network), the importance score of the nonlinear associated features is calculated along the preset time step. Based on the importance score, key associated features are selected from the nonlinear associated features, and the key associated features are retained in their original chronological order to form the initial feature time series. In this embodiment, through data preprocessing, local association capture, nonlinear transformation, and key selection, both data quality and core association information of meteorological and photovoltaic power output are accurately purified, providing a high-quality, high signal-to-noise ratio time series feature foundation for subsequent seasonal feature fusion and net load forecasting.

[0053] S2: Input the initial feature time series and the feature encoding representing the season into BiLSTM to obtain the enhanced feature time series that represents the correlation between meteorological factors and photovoltaic output, which represents the integrated seasonal characteristics.

[0054] The seasonal feature encoding adopts one-hot encoding, with spring being [1,0,0,0], summer being [0,1,0,0], autumn being [0,0,1,0], and winter being [0,0,0,1].

[0055] In some preferred embodiments, the step of inputting the initial feature time series sequence and the feature encoding representing the season into a BiLSTM to obtain the enhanced feature time series sequence representing the correlation between meteorological factors and photovoltaic output, which characterizes the fused seasonal characteristics, includes:

[0056] The initial feature time series is divided into a first initial feature time series and a second initial feature time series. Based on the seasonal characteristics and the first initial feature time series, the conventional single-path forget gate in BiLSTM is transformed into a seasonal enhanced forget gate. The second initial feature time series and the feature encoding are input into BiLSTM to obtain the enhanced feature time series.

[0057] Specifically, the modification of the conventional single-path forget gate in BiLSTM into a seasonally enhanced forget gate based on seasonal characteristics and the time-series sequence of the first initial features includes:

[0058] Based on the fluctuation trend of historical meteorological data in different seasons, meteorological thresholds for historical meteorological data in different seasons are obtained, and seasonal calibration items are obtained based on the meteorological thresholds.

[0059] A seasonal adaptation model is constructed based on seasonal calibration items and preset initial seasonal weights. The spliced ​​first initial feature time sequence and feature encoding are used as inputs to train the seasonal adaptation model to obtain the process seasonal weights.

[0060] If the process seasonal weights meet the preset conditions, the seasonal adaptation model is updated using the process seasonal weights. Based on the seasonal adaptation model, the conventional single-path forget gate is transformed into a seasonally enhanced forget gate. Otherwise, the process seasonal weights are used as the preset initial seasonal weights to obtain the seasonally enhanced forget gate.

[0061] The expression for obtaining the seasonal calibration term based on meteorological thresholds is: , The seasonal calibration term, representing season 's', is used to quantify the correlation between seasonal meteorological data fluctuations and photovoltaic power output. This represents the total number of time steps contained in season s. Indicates the number of categories of meteorological factors. This indicates the first step of time t under season s. Similar to historical meteorological data, Indicates the season s below the first The average meteorological threshold of historical meteorological data.

[0062] The expression for the seasonal adaptation model is:

[0063] , The process seasonal weight at time step t represents the strength of the influence of the current time step season on the meteorological-power output correlation. This represents the sigmoid activation function. This represents the trainable weight matrix of the seasonal adaptation model, and `concat` represents the concatenation operation. This represents the first initial feature time series sequence at time step t. The feature encoding for time step t, This represents the trainable bias term of the seasonal adaptation model. This indicates the initial seasonal weights preset based on experience. This represents the element-wise multiplication operation. When... When the seasonal weights of the process meet the preset conditions, This represents the seasonal weight of the process at time step t-1. This indicates the preset difference, which can be flexibly set according to user needs.

[0064] The expression for seasonally enhanced forgetfulness gates is:

[0065] ,in, The final weight of the seasonal enhancement forgetting gate is represented by time step t. The weights represent the weights of a standard single-path forget gate, and softmax represents the activation function. This represents the bias term for the seasonal weighted projection. pass Projecting onto the hidden layer dimension, then using softmax to generate attention weights, and modulating a regular single-path forget gate. The bias term representing the seasonal weighted projection; , This represents the trainable weight matrix of a conventional single-path forget gate. This represents the hidden state of the BiLSTM at the previous time step. This represents the feature at time step t in the first initial feature sequence. This represents the trainable bias term of a conventional single-path forget gate. In this embodiment, the conventional single-path forget gate in BiLSTM is transformed into a seasonally enhanced forget gate by using seasonal characteristics and a first initial feature time series sequence. This enables BiLSTM to obtain an enhanced feature time series sequence representing the correlation between meteorological factors and photovoltaic output, which are characterized by fused seasonal characteristics, from the spliced ​​second initial feature time series sequence and feature encoding. This solves the technical problem that existing technologies struggle to improve the accuracy of photovoltaic output prediction, resulting in low accuracy of net load prediction.

[0066] S3: The attention mechanism is used to process the enhanced feature time series to obtain the final weight of meteorological factors on photovoltaic output under different seasons.

[0067] In another embodiment, the step of processing the enhanced feature time series using an attention mechanism to obtain the final weights of meteorological factors for photovoltaic output in different seasons includes:

[0068] The seasonal coding vector in the enhanced feature time series is mapped to the query vector, the meteorological factor feature vector in the enhanced feature time series is mapped to the key vector, and the correlation feature vector between meteorological factors and photovoltaic output in the enhanced feature time series is mapped to the value vector. The similarity between the query vector and the key vector is calculated, and the final weight of meteorological factors for photovoltaic output in different seasons is obtained based on the similarity and the value vector.

[0069] Specifically, dot product attention is used to calculate the similarity between the query vector and the key vector. In this embodiment, the attention mechanism automatically learns the degree of influence of various meteorological factors on photovoltaic power output in all four seasons, and finally outputs the final weights of various meteorological factors in different seasons (e.g., radiation has the highest weight in winter, followed by temperature in autumn), thereby improving the accuracy of the weights of the influence of various meteorological factors on photovoltaic power output.

[0070] S4: Construct a weather forecasting model, use the model to obtain forecasted weather data, obtain forecasted photovoltaic (PV) output data based on the forecasted weather data and final weights, and obtain the forecasted net load based on the PV output data. The forecasted net load equals the total user load minus the forecasted PV output.

[0071] Specifically, the construction of a meteorological forecasting model and the use of the meteorological forecasting model to obtain forecast meteorological data include:

[0072] A weather forecasting model is constructed based on CNN and BiLSTM. Historical weather data is input into the weather forecasting model to obtain estimated weather data. The weather forecasting model is then corrected based on the difference between the estimated weather data and the actual weather data. Predicted weather data is then obtained based on the corrected weather forecasting model.

[0073] The process of obtaining photovoltaic power output forecast data based on predicted meteorological data and final weights includes:

[0074] The weight of the predicted meteorological data for photovoltaic power output is obtained based on the final weight, and the predicted photovoltaic power output data is obtained based on the weight of the predicted meteorological data for photovoltaic power output and the predicted meteorological data.

[0075] By constructing a weather forecasting model using CNN and BiLSTM, we retain the ability of CNN to capture local weather changes while leveraging BiLSTM to capture the long-term temporal dependencies of weather data. Combined with an estimation-correction mechanism, this makes the predicted weather data more closely reflect reality. When calculating photovoltaic power output, we no longer use a uniform, year-round weather weight; instead, we directly call the seasonally specific weather weights corresponding to the forecast period. This avoids the adaptation bias caused by the traditional one-size-fits-all approach to weights and effectively improves the accuracy of short-term net load forecasts.

[0076] In terms of overall prediction performance, the Attention-CNN-BiLSTM model significantly outperforms the other four comparative models across all four metrics and exhibits outstanding stability. For example... Figure 2 As shown, the average MAE of this invention across four seasons is as low as 0.336, a 91.3% reduction compared to CNN-BiLSTM (3.862), a 21.5% reduction compared to Attention-CNN-LSTM (0.428), a 22.0% reduction compared to Attention-BiLSTM (0.431), and a 22.2% reduction compared to LSTM (0.432); Figure 3 As shown, the average RMSE of this invention is 0.052, which is only 21.1% of LSTM (0.246) and 34.2% of Attention-BiLSTM (0.152), and the standard deviation of multi-site RMSE is only 0.018; Figure 4As shown, the average R² of this invention reaches 0.826, which is significantly higher than that of Attention-CNN-LSTM (0.457), Attention-BiLSTM (-0.218), CNN-BiLSTM (-0.583), and LSTM (-0.858); Figure 5 As shown, the average MAPE of this invention is 0.198, which is 30.5% lower than Attention-BiLSTM (0.285) and 36.5% lower than CNN-BiLSTM (0.312).

[0077] Example 2:

[0078] This embodiment also provides a short-term net load forecasting system that takes into account seasonal characteristics, applicable to the aforementioned short-term net load forecasting method that takes into account seasonal characteristics, including:

[0079] The initial feature time series acquisition module is used to extract features from historical photovoltaic power output data and historical meteorological data to obtain initial feature time series that characterize the correlation between meteorological factors and photovoltaic power output;

[0080] An enhanced feature time series acquisition module is used to input the initial feature time series and the feature encoding representing the season into BiLSTM to obtain an enhanced feature time series that represents the correlation between meteorological factors and photovoltaic output, which represents the fused seasonal characteristics.

[0081] The meteorological factor weight acquisition module is used to process the enhanced feature time series through an attention mechanism to obtain the final weight of meteorological factors for photovoltaic power output under different seasons.

[0082] The predicted net load acquisition module is used to construct a meteorological forecasting model, use the meteorological forecasting model to acquire predicted meteorological data, acquire photovoltaic predicted output data based on the predicted meteorological data and the final weights, and acquire the predicted net load based on the photovoltaic predicted output data.

[0083] Specifically, the initial feature time series sequence and the feature code representing the season are input into BiLSTM to obtain an enhanced feature time series sequence representing the correlation between meteorological factors and photovoltaic output that represent the integrated seasonal characteristics. This includes: dividing the initial feature time series sequence into a first initial feature time series sequence and a second initial feature time series sequence; modifying the conventional single-path forget gate in BiLSTM into a seasonal enhanced forget gate based on the seasonal characteristics and the first initial feature time series sequence; and inputting the second initial feature time series sequence and the feature code into BiLSTM to obtain an enhanced feature time series sequence. Based on seasonal characteristics and the first initial feature time series, the conventional single-path forget gate in BiLSTM is transformed into a seasonally enhanced forget gate. This includes: obtaining meteorological thresholds for historical meteorological data under different seasons based on the fluctuation trends of historical meteorological data under different seasons; obtaining seasonal calibration terms based on the meteorological thresholds; constructing a seasonal adaptation model based on the seasonal calibration terms and preset initial seasonal weights; training the seasonal adaptation model by using the spliced ​​first initial feature time series and feature encoding as inputs to obtain process seasonal weights; if the process seasonal weights meet preset conditions, updating the seasonal adaptation model through the process seasonal weights; and transforming the conventional single-path forget gate into a seasonally enhanced forget gate based on the seasonal adaptation model. Otherwise, using the process seasonal weights as preset initial seasonal weights, a seasonally enhanced forget gate is obtained. An attention mechanism is used to process the enhanced feature time series to obtain the final weights of meteorological factors on photovoltaic output in different seasons. This includes: mapping the seasonal encoding vector in the enhanced feature time series to a query vector, mapping the meteorological factor feature vector in the enhanced feature time series to a key vector, mapping the correlation feature vector between meteorological factors and photovoltaic output in the enhanced feature time series to a value vector, calculating the similarity between the query vector and the key vector, and obtaining the final weights of meteorological factors on photovoltaic output in different seasons based on the similarity and the value vector.

[0084] By modifying the conventional single-channel forget gate in BiLSTM into a seasonally enhanced forget gate using seasonal characteristics and the initial feature time series, the initial feature time series and the feature encoding representing the season are input into the seasonally enhanced forget gate to obtain an enhanced feature time series representing the correlation between meteorological factors and photovoltaic output, which integrates seasonal characteristics. Based on this, an attention mechanism is introduced to perform refined weighting processing on this enhanced sequence, adaptively assigning the final weights of each meteorological factor to photovoltaic output under different seasons. Finally, the predicted photovoltaic output data is obtained by combining the meteorological data of the prediction period with the final weight matrix, and the predicted net load is obtained based on the predicted photovoltaic output data. This solves the technical problem that existing technologies struggle to improve the accuracy of photovoltaic output prediction, leading to low accuracy in net load prediction.

[0085] Example 3:

[0086] This embodiment also provides a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the short-term net load forecasting method taking into account seasonal characteristics.

[0087] Example 4:

[0088] This embodiment also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the aforementioned short-term net load forecasting method taking into account seasonal characteristics.

[0089] The specific embodiments described above are preferred embodiments of a short-term net load forecasting method and system that takes into account seasonal characteristics, and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A short-term net load forecasting method considering seasonal characteristics, characterized by, The method comprises the following steps: characteristic extraction is performed on historical photovoltaic output data and historical meteorological data to obtain an initial feature time sequence representing the correlation between meteorological factors and photovoltaic output; the initial feature time sequence and a feature code representing seasons are input into a BiLSTM to obtain an enhanced feature time sequence representing the correlation between meteorological factors and photovoltaic output after fusing season characteristics; an attention mechanism is used to process the enhanced feature time sequence to obtain the most important weight of meteorological factors on photovoltaic output in different seasons; a meteorological prediction model is constructed, predicted meteorological data is obtained by using the meteorological prediction model, photovoltaic predicted output data is obtained based on the predicted meteorological data and the most important weight, and predicted net load is obtained based on the photovoltaic predicted output data; the initial feature time sequence and the feature code representing seasons are input into the BiLSTM to obtain the enhanced feature time sequence representing the correlation between meteorological factors and photovoltaic output after fusing season characteristics, which comprises: the initial feature time sequence is divided into a first initial feature time sequence and a second initial feature time sequence, a conventional single-path forgetting gate in the BiLSTM is modified into a season-enhanced forgetting gate based on season characteristics and the first initial feature time sequence, and the second initial feature time sequence and the feature code are input into the BiLSTM to obtain the enhanced feature time sequence; the conventional single-path forgetting gate in the BiLSTM is modified into the season-enhanced forgetting gate based on season characteristics and the first initial feature time sequence, which comprises: a meteorological threshold of historical meteorological data in different seasons is obtained according to fluctuation trends of the historical meteorological data in different seasons, and a season calibration term is obtained based on the meteorological threshold; a season adaptation model is constructed based on the season calibration term and a preset initial season weight, the first initial feature time sequence and the feature code are spliced to serve as input of the season adaptation model, the season adaptation model is trained to obtain a process season weight, and the season adaptation model is updated by the process season weight if the process season weight meets a preset condition, otherwise, the process season weight is taken as the preset initial season weight, so as to obtain the season-enhanced forgetting gate. characteristic extraction is performed on historical photovoltaic output data and historical meteorological data to obtain an initial feature time sequence representing the correlation between meteorological factors and photovoltaic output, which comprises:

2. The short-term net load forecasting method considering seasonal characteristics according to claim 1, characterized in that, time alignment processing and standardization processing are performed on the historical photovoltaic output data and the historical meteorological data to obtain an adaptive network structure; an initial correlation feature is obtained by capturing local correlation between meteorological factor changes and photovoltaic output changes through a convolution layer in a CNN along a preset time step on the adaptive network structure; nonlinear correlation features are obtained by performing nonlinear conversion on the initial correlation features through an activation function in the CNN, and key correlation features in the nonlinear correlation features are screened out along the preset time step, so as to obtain the initial feature time sequence. the most important weight of meteorological factors on photovoltaic output in different seasons is obtained by processing the enhanced feature time sequence through an attention mechanism, which comprises: 3.The short-term net load forecasting method considering seasonal characteristics according to claim 1, wherein, ​ The seasonal encoding vector in the enhanced feature time sequence is mapped as a query vector, the meteorological factor feature vector in the enhanced feature time sequence is mapped as a key vector, the association feature vector of the meteorological factor and the photovoltaic output in the enhanced feature time sequence is mapped as a value vector, and the similarity of the query vector and the key vector is calculated, and the most significant weight of the meteorological factor on the photovoltaic output in different seasons is obtained based on the similarity and the value vector.

4. The short-term net load forecasting method considering seasonal characteristics according to claim 1, characterized in that, The meteorological prediction model is constructed, and the predicted meteorological data is obtained by using the meteorological prediction model, which comprises: The meteorological prediction model is constructed based on the CNN and the BiLSTM, the estimated meteorological data is obtained by inputting the historical meteorological data into the meteorological prediction model, the meteorological prediction model is corrected according to the difference between the estimated meteorological data and the actual meteorological data, and the predicted meteorological data is obtained based on the corrected meteorological prediction model.

5. The short-term net load forecasting method considering seasonal characteristics according to claim 1, characterized in that, The photovoltaic predicted output data is obtained based on the predicted meteorological data and the most significant weight, which comprises: The weight of the predicted meteorological data on the photovoltaic output is obtained according to the most significant weight, and the photovoltaic predicted output data is obtained based on the weight of the predicted meteorological data on the photovoltaic output and the predicted meteorological data.

6. A short-term net load forecasting system considering seasonal characteristics, which is suitable for a short-term net load forecasting method considering seasonal characteristics according to any one of claims 1 to 5, characterized in that, It comprises: An initial feature time sequence acquisition module is configured to extract features from the photovoltaic historical output data and the historical meteorological data to obtain an initial feature time sequence representing the association between the meteorological factors and the photovoltaic output; An enhanced feature time sequence acquisition module is configured to input the initial feature time sequence and the feature encoding representing the season into the BiLSTM to obtain an enhanced feature time sequence representing the association between the meteorological factors and the photovoltaic output with the fusion of the seasonal characteristics; A meteorological factor weight acquisition module is configured to process the enhanced feature time sequence by the attention mechanism to obtain the most significant weight of the meteorological factor on the photovoltaic output in different seasons; A predicted net load acquisition module is configured to construct a meteorological prediction model, obtain predicted meteorological data by using the meteorological prediction model, obtain photovoltaic predicted output data based on the predicted meteorological data and the most significant weight, and obtain predicted net load based on the photovoltaic predicted output data.

7. A computer device, characterized by: It comprises: A processor, a memory and a bus, the memory stores machine readable instructions executable by the processor, when the computer equipment is running, the processor and the memory communicate through the bus, the processor executes the machine readable instructions to execute the steps of the short-term net load prediction method considering seasonal characteristics in any one of claims 1-5.

8. A computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is executed by a processor to execute the steps of the short-term net load prediction method considering seasonal characteristics in any one of claims 1-5.

Images