A power selling company load prediction method based on a prophet model
By constructing the Prophet model through a step-by-step optimization strategy, the problem of insufficient load forecasting accuracy for high-energy-consuming enterprises was solved, the forecasting accuracy and robustness were improved, and the model was adapted to the load characteristics of large industrial users, especially for shutdown and production stoppage events, thus achieving complete forecasting results and risk management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINGDEZHEN POWER PLANT OF STATE POWER INVESTMENT GRP JIANGXI ELECTRIC POWER CO LTD
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-14
AI Technical Summary
Existing load forecasting methods based on the Prophet model suffer from insufficient forecasting accuracy and poor robustness in high-energy-consuming enterprises, making it difficult to adapt to the load characteristics of large industrial users. Furthermore, they do not include production events such as shutdowns and stoppages as independent influencing factors in the model.
A step-by-step optimization strategy was adopted to construct a load forecasting model, which sequentially solved for the trend term, seasonal term, and special date impact term, determined the trend, cycle, and event disturbances respectively, included the shutdown and production stoppage events of high energy-consuming enterprises as independent factors, and constructed a probabilistic forecasting interval.
It improves the stability and interpretability of the model, significantly reduces the prediction bias of key nodes, provides point prediction values and risk ranges, and meets the needs of power purchase planning and power transaction risk management.
Smart Images
Figure CN122390326A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of load forecasting, and in particular relates to a load forecasting method for electricity sales companies based on the Prophet model. Background Technology
[0002] As a key player in the electricity market, electricity sales companies are responsible for purchasing electricity in bulk from power generation companies or electricity trading centers and selling it to end users. Load forecasting is a fundamental task for electricity sales companies in conducting electricity trading, formulating power purchase plans, and ensuring users' electricity needs. The accuracy of forecasting directly affects the stable operation of the power system and the efficiency of power resource allocation.
[0003] For electricity sales companies that primarily serve energy-intensive enterprises, load forecasting is of paramount importance. Taking the electrolytic aluminum industry as an example, electrolytic aluminum is a key supported and developed basic raw material industry. During production, electrolytic cells require a continuous and stable power supply. Sudden power outages or insufficient power supply can lead to cell solidification, equipment damage, and even safety accidents. Therefore, accurate load forecasting by electricity sales companies, allowing them to secure sufficient power in the electricity market in advance, is crucial for ensuring the continuous and stable production of energy-intensive enterprises. Furthermore, accurate load forecasting allows power grid dispatching agencies to make appropriate generation plans and transmission arrangements, avoiding problems such as power system frequency fluctuations and voltage instability caused by load fluctuations. This ensures the safety and stability of power supply and helps electricity sales companies optimize their power purchase strategies, rationally allocating power purchase ratios in the spot and medium-to-long-term markets, reducing power purchase costs, and improving the efficiency of power resource allocation.
[0004] Currently, load forecasting methods based on the Prophet algorithm have been applied in the power sector. Existing technical solutions typically employ joint optimization, simultaneously solving for all parameters of the trend, seasonal, and holiday terms, with the model focusing only on minimizing the error between the final predicted value and the actual value. However, such solutions still have the following technical problems in practical applications: First, joint optimization may cause the model to distribute errors among the decomposition terms in order to fit the training data, resulting in the trend term absorbing periodic fluctuations that should be explained by the seasonal term, or the seasonal term absorbing special event disturbances that should be explained by the holiday term, leading to feature confusion. Second, existing solutions typically only consider the impact of public holidays on the load. For power sales companies supplying electricity to high-energy-consuming enterprises, shutdowns and production stoppages are the most influential load disturbance events, but existing solutions do not include them as independent influencing factors in the model. Summary of the Invention
[0005] This invention provides a load forecasting method for electricity sales companies based on the Prophet model, which solves the problems of insufficient forecasting accuracy, poor robustness, and difficulty in adapting to the load characteristics of large industrial users in existing load forecasting methods.
[0006] The basic solution provided by this invention is a load forecasting method for electricity sales companies based on the Prophet model, comprising the following steps: S1: Collect historical load data from the electricity sales company. The historical load data includes load time series data and external factor data that affect load changes. The external factor data includes meteorological factor data, calendar factor data, and production event data. The collected data is then preprocessed. S2: Construct and train a load forecasting model based on the Prophet algorithm. The load forecasting model includes a trend term, a seasonal term, a special date impact term, and an error term. When training the model, a step-by-step optimization strategy is used to train the load forecasting model. S3: Input the preprocessed historical load data into the trained load prediction model to make a preliminary prediction of the electricity load of the electricity sales company's agents and obtain a deterministic prediction value. S4: Based on the distribution characteristics of the residual sequence generated during the training process, combined with the deterministic predicted values, a probabilistic prediction interval is constructed to obtain the final prediction result of the electricity sales company's load. The final prediction result includes point predicted values and corresponding confidence intervals.
[0007] Preferably, the step-by-step optimization strategy includes: a. Input historical load data into the trend term model, and use a saturated growth model to determine the trend term. The saturated growth model is set according to the industry characteristics of the electricity sales company's load being limited by the capacity ceiling. b. Using the trend term determined in step a as a known quantity, use Fourier series to fit the residual sequence after deducting the trend term from the historical load data to determine the seasonal term containing daily, weekly, and monthly cycles. The parameters of the trend term remain unchanged during the fitting process. c. Using the trend and seasonal terms determined in steps a and b as known quantities, extract the occurrence time and duration of all special date events based on the calendar factor data and shutdown / production stoppage information in the historical load data. Establish a time window function to fit the residual sequence after deducting the trend and seasonal terms from the historical load data, determine the influence coefficient corresponding to each shutdown event window, and obtain the special date influence term. During the fitting process, the parameters of the trend and seasonal terms remain unchanged, and the final residual sequence is obtained. The special date events include holiday events corresponding to the calendar factor data and shutdown / production stoppage events corresponding to the production event data.
[0008] More preferably, the saturated growth model is expressed as:
[0009] Where k represents the average periodic growth rate of the load over time t, δ is the change in the load growth rate, and m is the offset. This indicates the indicator function, where γ represents the offset adjustment amount, and , For the practice of the j-th variable point, This indicates the maximum load capacity.
[0010] More preferably, in the step-by-step optimization strategy, the seasonal term in step b... Constructed using Fourier series, it can be represented as:
[0011] in, Here, is the seasonal function representing the periodic fluctuation component of the load; P represents the period of the time series, corresponding to days, weeks, or months; and N is the Fourier series order, used to control the complexity of the seasonal term fitting. This represents the cosine coefficient of the nth harmonic. The coefficients of the sinusoidal term of the nth harmonic are... This represents the angular frequency of the nth harmonic.
[0012] More preferably, in the step-by-step optimization strategy, in step c, the expression for the time window function is:
[0013]
[0014]
[0015] in, This represents the impact of a special date, where i represents the i-th special date event, L represents the total number of events, and Di represents the time range affected by the i-th special date event, which is set based on the event's occurrence time and duration. Let ki represent the indicator function, where ki is the magnitude of the impact of the i-th specific date event on the load, determined by fitting the residual sequence. The indicator vector is a row vector consisting of L indicator functions, representing the activation state of each event window at time t. The coefficient vector is a column vector consisting of L influence coefficients. The time variable is t, i.e., the current calculation time.
[0016] Preferably, in step S1, the preprocessing includes: For abnormal data in historical load data, a range of variation is set based on the load data at adjacent time points, and load data that exceeds this range is replaced with the average value of the load data at adjacent time points. For missing data in historical load data, the average value of load data at the preceding and following time points is used to fill in the missing data.
[0017] Preferably, it also includes S5: After the model training is completed, the test set data is input into the trained load prediction model, and the error index is calculated based on the trend term prediction value, the trend and seasonal superposition prediction, and the complete deterministic prediction value, respectively. By comparing the error index of the prediction values at each stage, the contribution of the trend term, seasonal term, and special date term to the prediction accuracy is quantified.
[0018] More preferably, the prediction error is calculated using at least one of mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).
[0019] The principles and advantages of this invention are as follows: 1. A step-by-step optimization strategy is adopted to solve the trend term, seasonal term, and special date influence term in sequence. This avoids the mutual absorption of errors and confusion of features by each decomposition term due to joint optimization, making the physical meaning of trends, cycles, and event disturbances clearer and significantly improving the stability and interpretability of the model.
[0020] 2. Production events such as shutdowns, production stoppages, and maintenance of high-energy-consuming enterprises are included in the model as independent special date items, which makes up for the shortcomings of existing technologies that only consider public holidays, better reflects the sudden change pattern of industrial users' load, and significantly reduces the prediction bias of key nodes.
[0021] 3. Based on point forecasts, confidence intervals are constructed based on residual distributions to provide electricity sales companies with complete results of point forecast values and risk intervals. This not only meets the needs of electricity purchase plan preparation but also provides quantitative basis for power transaction risk management and reserve capacity allocation. Attached Figure Description
[0022] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0023] The following detailed description illustrates the specific implementation method: The specific implementation process is as follows: (See details) Figure 1 A load forecasting method for electricity retailers based on the Prophet model includes the following steps: S1: Collect historical load data from the electricity sales company. The historical load data includes load time series data and external factor data that affect load changes. The external factor data includes meteorological factor data, calendar factor data, and production event data. The collected data is then preprocessed.
[0024] Among them, load time series data is obtained from the power marketing system or grid dispatch system of the electricity sales company, recording the active power value at each time point; meteorological factor data is obtained from the local meteorological department or meteorological data service platform, including temperature, humidity, wind speed, etc.; calendar factor data is generated according to the national statutory holiday arrangement, including holiday marking information; production event data is obtained from the production plans of industrial enterprises served by the electricity sales company, including at least shutdown and production stoppage marking information.
[0025] In step S1, the preprocessing includes: The data is cleaned to remove erroneous and duplicate values and fill in missing values. Specifically, for abnormal data in historical load data, a range of variation is set based on the load data at adjacent time points, and load data exceeding this range is replaced with the average value of the load data at adjacent time points; for missing data in historical load data, the average value of the load data at the preceding and following time points is used to fill in the missing data.
[0026] It also includes dividing the preprocessed data into training and testing sets for subsequent model training and evaluation.
[0027] In this embodiment, historical load data from a power sales company from January 1, 2019 to December 31, 2021 was collected. The sampling frequency was once per hour, resulting in a total of 26,280 load data points.
[0028] Simultaneously, data on external factors affecting load changes are collected: Meteorological data: Temperature (°C) and humidity (%) at the corresponding time points are from local weather station records; Calendar factor data: Holiday information, marking national statutory holidays (New Year's Day, Spring Festival, Qingming Festival, Labor Day, Dragon Boat Festival, Mid-Autumn Festival, National Day) as 1, and non-holidays as 0; Production event data: Stoppage and shutdown information. Based on the annual maintenance plan and temporary shutdown records of the electrolytic aluminum enterprises served by the power sales company, the stoppage and shutdown days are marked as 1, and normal production days are marked as 0. The collected data was cleaned, and abnormal and missing data were processed. Specifically, when processing abnormal data, the load data of two time points before and after the current time point were used as a benchmark, and the variation range was set to ±20%. When the load data at a certain time point exceeded this range, it was determined to be abnormal data, and the average value of the load data of the two time points before and after was used to replace it. In this embodiment, a total of 127 abnormal data points were identified and processed. When processing missing data, for data missing due to instrument failure or communication interruption, the average value of the load data of two time points before and after the missing time point was used to fill it. In this embodiment, a total of 83 missing data points were processed. The preprocessed data was then divided into training and test sets in an 8:2 ratio.
[0029] S2: Construct and train a load forecasting model based on the Prophet algorithm, wherein the load forecasting model includes a trend term. Seasonal items Special Dates Affecting Items Error Term During model training, a step-by-step optimization strategy is used to train the load prediction model.
[0030] The load forecasting model is expressed as follows:
[0031] in, This represents the load forecast value. Describe the trend item. Indicates seasonal items, Indicates items affected by special dates. Indicates the error term; In step S2, the step-by-step optimization strategy includes: a. Input historical load data into the trend term model and use a saturated growth model to determine the trend term. The saturated growth model is set according to the industry characteristics of electricity sales companies where the load is limited by the upper limit of production capacity; The saturated growth model is expressed as:
[0032] Where k represents the average periodic growth rate of the load over time t, δ is the change in the load growth rate, and m is the offset. This indicates the indicator function, where γ represents the offset adjustment amount, and , For the practice of the j-th variable point, This indicates the maximum load capacity.
[0033] In this embodiment, three points of change in growth rate—June 2019, March 2020, and November 2020—were automatically identified, and a trend term was obtained through fitting. After fitting, the first step of calculating the residual sequence is performed. ,in, This represents the true historical load value, which is the load time series data collected and preprocessed in step S1.
[0034] b. The trend term determined in step a. Given the quantities, use Fourier series to subtract the trend term from the historical load data. The residual sequence is then fitted to determine the seasonal terms that include daily, weekly, and monthly cycles. Trend term during fitting process The parameters remain unchanged; The seasonal items Constructed using Fourier series, it can be represented as:
[0035] in, Here, is the seasonal function representing the periodic fluctuation component of the load; P represents the period of the time series, corresponding to days, weeks, or months; and N is the Fourier series order, used to control the complexity of the seasonal term fitting. This represents the cosine coefficient of the nth harmonic. The coefficients of the sinusoidal term of the nth harmonic are... This represents the angular frequency of the nth harmonic.
[0036] Specifically, different parameters are set for different periods: Daily cycle: P=1, N=5, fitting the 24-hour load fluctuation pattern within a day; Weekly cycle: P=7, N=3, fitting the load difference between weekdays and weekends; Monthly cycle: P=30.44, N=2, fitting the monthly load fluctuation pattern; The Fourier coefficients were obtained by fitting using the least squares method. , Determine the seasonal items After fitting is complete, the residual sequence for the second step is calculated. .
[0037] c. Trend terms determined by steps a and b and seasonal items Given the historical load data, based on calendar factor data and shutdown / production stoppage indicators, extract the occurrence time and duration of all special date events. Establish a time window function to subtract trend terms from the historical load data. and seasonal items The residual sequence is then fitted to determine the impact coefficient corresponding to each shutdown event window, thus obtaining the impact item for special dates. Trend term during fitting process and seasonal items The parameters remain unchanged to obtain the final residual sequence, wherein the special date events include holiday events corresponding to calendar factor data and shutdown events corresponding to production event data.
[0038] In step c, the expression for the time window function is:
[0039]
[0040]
[0041] in, This represents the impact of a special date, where i represents the i-th special date event, L represents the total number of events, and Di represents the time range affected by the i-th special date event, which is set based on the event's occurrence time and duration. Let ki represent the indicator function, where ki is the magnitude of the impact of the i-th specific date event on the load, determined by fitting the residual sequence. The indicator vector is a row vector consisting of L indicator functions, representing the activation state of each event window at time t. The coefficient vector is a column vector consisting of L influence coefficients. The time variable is t, i.e., the current calculation time.
[0042] In this embodiment, a total of 42 special date events were identified, including 36 national statutory holidays and 6 enterprise shutdown and production stoppage events. A time window was defined for each event. Set different window lengths based on event type: Holiday events: Set the window to [1 day before the holiday, the day of the holiday, 1 day after the holiday]; Stoppage / Production Stoppage Events: Set the window to [2 days before stoppage, during stoppage, 2 days after resumption of work] to cover the entire process of equipment shutdown preparation and production resumption.
[0043] After fitting, the final residual sequence is obtained. This is used for subsequent prediction interval construction.
[0044] S3: Input the preprocessed historical load data into the trained load prediction model to make a preliminary prediction of the electricity load of the electricity sales company's agents and obtain a deterministic prediction value. By inputting preprocessed historical load data into the trained load forecasting model, a deterministic forecast value for January 2022 is obtained. In this embodiment, taking 0:00 on January 1, 2022 as an example, the deterministic prediction value for that time point is calculated to be 312.5MW.
[0045] S4: Based on the distribution characteristics of the residual sequence generated during the training process, combined with the deterministic predicted values, a probabilistic prediction interval is constructed to obtain the final prediction result of the electricity sales company's load. The final prediction result includes point predicted values and corresponding confidence intervals.
[0046] Based on the final residual sequence obtained in step c Based on the distribution characteristics, the standard deviation σ of the residuals is estimated. In this embodiment, σ = 8.3MW is calculated.
[0047] Based on deterministic predicted values Using the standard deviation σ, construct a prediction interval with a 95% confidence level: Upper Realm: + 1.96 × σ = 312.5 + 1.96 × 8.3 = 328.8MW; The lower realm: - 1.96 × σ = 312.5 - 1.96 × 8.3 = 296.2MW; The final prediction result is as follows: The point prediction value at 0:00 on January 1, 2022 is 312.5MW, with a 95% confidence interval of [296.2MW, 328.8MW].
[0048] It also includes S5: After the model training is completed, the test set data is input into the trained load forecasting model, and the error index is calculated based on the trend term forecast value, the trend and seasonal superposition forecast, and the complete deterministic forecast value. By comparing the error index of the forecast values at each stage, the contribution of the trend term, seasonal term, and special date term to the forecast accuracy is quantified.
[0049] The prediction error is calculated using at least one of the following: mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).
[0050] Specifically, the mean squared error (MSE) is expressed as follows:
[0051] Where n represents the total number of samples, For the true value, The predicted value is MSE; MSE can evaluate the degree of variation in the data. It equals 0 when the predicted value perfectly matches the actual value, which is the perfect model; the larger the error, the larger the value.
[0052] The root mean square error (RMSE) is expressed as follows:
[0053] Where n represents the total number of samples, For the true value, The predicted value is RMSE. RMSE can be used to measure the deviation between the observed value and the true value. Its significance lies in the fact that after taking the square root, the result of the error is at the same level as the data, which can better describe the data. For example, RMSE=10 can be considered as the regression effect differing from the true value by an average of 10. When the predicted value and the true value are completely in agreement, the value is equal to 0, which is the perfect model. The larger the error, the larger the value.
[0054] The expression for Mean Absolute Error (MAE) is:
[0055] Where n represents the total number of samples, For the true value, The predicted value is denoted as MAE. MAE can better reflect the actual situation of the prediction error. When the predicted value matches the true value perfectly, it equals 0, which is the perfect model. The larger the error, the larger this value is.
[0056] The expression for Mean Absolute Percentage Error (MAPE) is:
[0057] Where n represents the total number of samples, For the true value, The value is the predicted value; a MAPE of 0% indicates a perfect model, while a MAPE greater than 100% indicates a poor model. The smaller the MAPE value, the better the accuracy of the prediction model.
[0058] After the model training was completed, the test set data (672 time points) from February 2022 was input into the trained load forecasting model, and the mean absolute error (MAE) of the three forecast values was calculated respectively: Based solely on trend items Prediction: = 28.7MW Based on trend and seasonal items + Prediction: = 15.3MW Based on the complete model + + Prediction: = 8.9MW By comparing the error indices at each stage, the contribution of each decomposition term to the prediction accuracy is quantified: Seasonal contributions: - = 13.4MW, meaning the seasonal term reduced the forecast error by 46.7%. Special date item contribution: - = 6.4MW, meaning the special date factor further reduced the forecast error by 41.8%; The results show that both seasonal and special date terms contribute significantly to improving prediction accuracy, with the seasonal term contributing the most and the special date term playing a key role during holidays and shutdowns.
[0059] Example 2 The difference from Example 1 is that, in the step-by-step optimization strategy, holiday events and work stoppage / production shutdown events are handled separately: Specifically, step c includes: c1. Trend term determined by steps a and b and seasonal items Given the known quantities, based on the calendar factor data in the historical load data, the occurrence time and duration of each holiday event are extracted. A time window is defined for each event, and the holiday item logic of the Prophet algorithm is used to deduct the trend item from the historical load data. and seasonal items The residual sequence is then fitted to determine the influence coefficient corresponding to each holiday event window. Received items affected by holidays Trend term during fitting process and seasonal items The parameters remain unchanged; c2. Trend terms determined by steps a, b, and c1 Seasonal items and the impact of holidays Given the known quantities, based on the shutdown and production stoppage indicators in the historical load data, the occurrence time and duration of each shutdown and production stoppage event are extracted. A time window is defined for each event, and the holiday item logic of the Prophet algorithm is used to deduct trend items from the historical load data. Seasonal items By fitting the residual sequence after the holiday impact term h_holiday(t), the impact coefficient corresponding to each shutdown event window is determined. Received the impact of work stoppage and production shutdown Trend term during fitting process Seasonal items and the impact of holidays The parameters remain unchanged, and the final residual sequence is obtained; Among them, special date impact items .
[0060] Separating holiday events and work stoppage / production stoppage events from the impact items of special dates can further avoid mutual interference between holiday events and work stoppage / production stoppage events. This is suitable for scenarios where there are many holiday events and work stoppage / production stoppage events or where the impact patterns of the two types of events are significantly different.
[0061] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A load forecasting method for electricity sales companies based on the Prophet model, characterized in that, Includes the following steps: S1: Collect historical load data from the electricity sales company. The historical load data includes load time series data and external factor data that affect load changes. The external factor data includes meteorological factor data, calendar factor data, and production event data. The collected data is then preprocessed. S2: Construct and train a load forecasting model based on the Prophet algorithm. The load forecasting model includes a trend term, a seasonal term, a special date impact term, and an error term. When training the model, a step-by-step optimization strategy is used to train the load forecasting model. S3: Input the preprocessed historical load data into the trained load prediction model to make a preliminary prediction of the electricity load of the electricity sales company's agents and obtain a deterministic prediction value. S4: Based on the distribution characteristics of the residual sequence generated during the training process, combined with the deterministic predicted values, a probabilistic prediction interval is constructed to obtain the final prediction result of the electricity sales company's load. The final prediction result includes point predicted values and corresponding confidence intervals.
2. The load forecasting method for electricity sales companies based on the Prophet model according to claim 1, characterized in that: In step S2, the step-by-step optimization strategy includes: a. Input historical load data into the trend term model, and use a saturated growth model to determine the trend term. The saturated growth model is set according to the industry characteristics of the electricity sales company's load being limited by the capacity ceiling. b. Using the trend term determined in step a as a known quantity, use Fourier series to fit the residual sequence after deducting the trend term from the historical load data to determine the seasonal term containing daily, weekly, and monthly cycles. The parameters of the trend term remain unchanged during the fitting process. c. Using the trend and seasonal terms determined in steps a and b as known quantities, extract the occurrence time and duration of all special date events based on the calendar factor data and shutdown / production stoppage information in the historical load data. Establish a time window function to fit the residual sequence after deducting the trend and seasonal terms from the historical load data, determine the influence coefficient corresponding to each shutdown event window, and obtain the special date influence term. During the fitting process, the parameters of the trend and seasonal terms remain unchanged, and the final residual sequence is obtained. The special date events include holiday events corresponding to the calendar factor data and shutdown / production stoppage events corresponding to the production event data.
3. The load forecasting method for electricity sales companies based on the Prophet model according to claim 2, characterized in that: The saturated growth model is expressed as: in, This represents the trend term, where k represents the average periodic growth rate of load change over time t, δ is the change in load growth rate, and m is the offset. This indicates the indicator function, where γ represents the offset adjustment amount, and , For the practice of the j-th variable point, This indicates the maximum load capacity.
4. The load forecasting method for electricity sales companies based on the Prophet model according to claim 2, characterized in that: In the step-by-step optimization strategy, the seasonal term in step b... Constructed using Fourier series, it can be represented as: in, Here, is the seasonal function representing the periodic fluctuation component of the load; P represents the period of the time series, corresponding to days, weeks, or months; and N is the Fourier series order, used to control the complexity of the seasonal term fitting. This represents the cosine coefficient of the nth harmonic. The coefficients of the sinusoidal term of the nth harmonic are... This represents the angular frequency of the nth harmonic.
5. The load forecasting method for electricity sales companies based on the Prophet model according to claim 2, characterized in that: In the step-by-step optimization strategy, the time window function expression in step c is: in, This represents the impact of a special date, where i represents the i-th special date event, L represents the total number of events, and Di represents the time range affected by the i-th special date event, which is set based on the event's occurrence time and duration. Let ki represent the indicator function, where ki is the magnitude of the impact of the i-th specific date event on the load, determined by fitting the residual sequence. The indicator vector is a row vector consisting of L indicator functions, representing the activation state of each event window at time t. The coefficient vector is a column vector consisting of L influence coefficients. The time variable is t, i.e., the current calculation time.
6. The load forecasting method for electricity sales companies based on the Prophet model according to claim 1, characterized in that: In step S1, the preprocessing includes: For abnormal data in historical load data, a range of variation is set based on the load data at adjacent time points, and load data that exceeds this range is replaced with the average value of the load data at adjacent time points. For missing data in historical load data, the average value of load data at the preceding and following time points is used to fill in the missing data.
7. The load forecasting method for electricity sales companies based on the Prophet model according to claim 1, characterized in that: It also includes S5: After the model training is completed, the test set data is input into the trained load forecasting model, and the error index is calculated based on the trend term forecast value, the trend and seasonal superposition forecast, and the complete deterministic forecast value. By comparing the error index of the forecast values at each stage, the contribution of the trend term, seasonal term, and special date term to the forecast accuracy is quantified.
8. The load forecasting method for electricity sales companies based on the Prophet model according to claim 7, characterized in that: The prediction error is calculated using at least one of the following: mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).