Thermal power plant demand prediction method based on holt-winters

Abstract

The invention discloses a thermal power plant demand prediction method based on holt-winters. The method comprises the steps that S1, a sales settlement data matrix in a preset historical period is acquired, rows in the matrix represent years, and columns in the matrix represent months; s2, respectively substituting the historical data in the sales settlement data matrix into a holt-winters addition model and a multiplication model; s3, traversing all historical data in the sales settlement data matrix, continuously fitting values of three hyper-parameters including a horizontal item, a trend item and a seasonal item in the holt-winters model, and determining an optimal parameter combination according to the hyper-parameter corresponding to the minimum average absolute percentage error in the model fitting process; and S4, predicting the electric quantity demand and the heat demand of the target thermal power plant in a future preset month according to the optimal combination. According to the prediction method, the trend and seasonal characteristics of the electric heating demand are captured at the same time, the limitation of single historical data reference in the same period is solved, and reliable data support is provided for production scheduling plan making.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of thermal power plant demand prediction, in particular to a thermal power plant demand prediction method based on holt-winters. BACKGROUND

[0002] In the face of increasingly fierce market competition, the sales strategy of thermal power enterprises, especially of combined heat and power enterprises, becomes a crucial link for their future profitability, and the demand prediction is the most important part of the production and sales strategy.

[0003] In the prior art, the sales volume of heat is often simply referred to user demand or completely depends on the subjective judgment or individual opinion of the user side manager, resulting in a lack of objective analysis in sales volume confirmation; the power prediction adopts the traditional naive prediction method, which is only applicable to special periods in the peak season, and in the off-season, the prediction accuracy is extremely low due to lack of data support. SUMMARY

[0004] The present application provides a thermal power plant demand prediction method based on holt-winters, which adopts time series prediction means to solve the demand prediction problem of combined heat and power enterprises, and is mainly used for predicting the demand of electricity and heat to help solve the key problems of production scheduling and sales allocation.

[0005] The embodiment of the present application provides a thermal power plant demand prediction method based on holt-winters, which comprises:

[0006] S1, obtaining a sales settlement data matrix in a preset historical period, wherein the rows in the matrix represent years and the columns represent months;

[0007] S2, bringing the historical data in the sales settlement data matrix into a holt-winters additive model and a holt-winters multiplicative model respectively;

[0008] S3, traversing all the historical data in the sales settlement data matrix, constantly fitting the values of three hyperparameters of a holt-winters model, i.e., a horizontal term, a trend term and a seasonal term, and determining the best parameter combination according to the hyperparameters corresponding to the minimum mean absolute percentage error in the model fitting process;

[0009] S4, predicting the electricity demand and heat demand of the target thermal power plant in the future preset months according to the best combination.

[0010] Optionally, the S3 comprises:

[0011] When the three hyperparameters are in the range of (0, 1), the first preset step size 0.1 is taken as the search step size, and the minimum mean absolute percentage error value searched is taken as the first best combination parameter;

[0012] Centered on the hyperparameters corresponding to the first optimal combination, a traversal search is performed within the range of ±0.1 with a second predetermined step size of 0.01, and the minimum mean absolute percentage error value obtained from the search is taken as the second optimal combination parameter.

[0013] Centered on the hyperparameters corresponding to the second optimal combination, a traversal search is performed within the range of ±0.01 with a third predetermined step size of 0.001, and the minimum mean absolute percentage error value obtained from the search is taken as the final optimal parameter combination.

[0014] Optionally, the sales settlement data includes the demand in the electricity market and the heat market, wherein the demand in the electricity market consists of the annual transaction volume, the market transaction volume before the month, and the market transaction volume within the month.

[0015] Optionally, for electricity market demand, when planned shutdowns occur, a moving average of previous and subsequent data is used to adjust the monthly demand. The adjustment formula is as follows:

[0016]

[0017] In the formula, This is for the correction requirements in month i.

[0018] Optionally, for electricity market demand, when unplanned shutdowns occur, an interpolation correction is applied based on the ratio of normal working days to abnormal working days. The correction formula is as follows:

[0019]

[0020] In the formula, This represents the number of normal working days for the unit in month i, calculated in calendar days per month.

[0021] This represents the number of abnormal days for the unit in month i.

[0022] Optionally, the electricity demand data and heat demand data can be fitted into a virtual steam market demand data using thermodynamic principles. The fitting method is as follows:

[0023]

[0024] C H,S ~C H h H

[0025] In the formula, C E Let C be the demand parameter for a given electricity market. H For the demand parameters of a certain heat market;

[0026] C E,SThe steam equivalent value converted from the demand parameters of the electricity market;

[0027] C H,S The steam equivalent value converted from the demand parameters of the heat market;

[0028] d0 is the power generation steam consumption rate;

[0029] h H Specific enthalpy of heating steam.

[0030] Optionally, after S4, the method further includes:

[0031] Plot a comparison graph of the original data, historical fitted values, and future predicted values, and use different colors and line styles to distinguish the results of the additive model and the multiplicative model.

[0032] The beneficial effects of this invention are:

[0033] 1. Employing a time series forecasting framework, compared to the limitations of traditional naive forecasting methods that only consider historical data from the same period, it can simultaneously capture both the long-term trend and seasonal fluctuations in electricity and heat demand, significantly improving forecast accuracy and reducing decision-making risk; Adaptive model selection: Through dual-path calculation using additive and multiplicative models, combined with 39706 parameter combination iterations (11 3 +21 3 +21 3 This enables adaptive matching of different data characteristics, solving the problem that a single model cannot cope with complex electrothermal load characteristics.

[0034] 2. A shutdown data repair mechanism was set up, and different repair methods were adopted for planned or unplanned shutdowns, which effectively eliminated the interference of abnormal unit operating conditions on historical data and improved the accuracy of basic forecast data; the organic integration of annual contract, pre-month market and intra-month market electricity data was realized, which solved the problem of "conservative at the beginning of the year and rush at the end of the year" caused by the fragmentation of market data in traditional forecasting.

[0035] 3. By using thermodynamic principles, the electricity and heat demand are converted into unified steam equivalent data, providing a quantitative basis for fuel allocation of cogeneration units; comparative charts of raw data, historical fitted values ​​and predicted values ​​for the next 12 months are generated, and the visualization output supports intuitive comparison of results from different models, helping management to quickly identify forecast trends.

[0036] 4. A three-level progressive strategy of "low-precision grid search (step size 0.1) → medium-precision local optimization (step size 0.01) → high-precision fine-tuning (step size 0.001)" is adopted to ensure prediction accuracy while controlling the calculation time within 50.88 seconds, meeting the timeliness requirements of the company's monthly forecast. Attached Figure Description

[0037] Figure 1 A flowchart of demand forecasting for thermal power plants based on Holt-Winters is provided for embodiments of the present invention;

[0038] Figure 2 For the statistical table of unit shutdown status of Company X from 2020 to 2024 provided for the embodiments of the present invention, please refer to [link / reference]. Figure 2 ;

[0039] Figure 3 This is a power demand data chart provided in an embodiment of the present invention;

[0040] Figure 4 This is a chart of heat demand data provided in an embodiment of the present invention;

[0041] Figure 5 This is a chart of steam market demand data provided in an embodiment of the present invention. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0043] This invention provides a method for demand forecasting of thermal power plants based on Holt-Winters, a forecasting algorithm for processing time series data with trend and seasonal characteristics. This method uses exponential smoothing to weight historical data, extracts trend and seasonal components, and thus predicts future trends.

[0044] 1. Feasibility analysis and method selection for time series forecasting requirements

[0045] Information asymmetry is inevitable among participants in the electricity market. Because competitors' trading data is highly commercially sensitive, market participants often face difficulties in acquiring such data. In this context, using a company's own historical trading data for predictive modeling may be the only feasible approach, but this requires feasibility analysis.

[0046] Electricity and heat demand exhibits significant time sensitivity, primarily manifested in seasonality and trends. To achieve relatively accurate demand forecasting, the most suitable time series forecasting method should be used. Considering the practical situation, for example, Company X's existing data consists of a set of sales data spanning 5 years × 12 months, which is relatively scarce. Therefore, the model used in this invention must be able to produce relatively reliable results even with limited metadata and potentially high noise levels. With the development of machine learning algorithms, advanced algorithms such as Neural Networks (NN) and Artificial Neural Networks (ANN) are gradually emerging in data forecasting. However, these algorithms all require a sufficient amount of data for analysis, which is clearly insufficient for the machine learning algorithms described in this invention. Based on this, this invention summarizes existing research results across various industries and, referencing various literature, summarizes the characteristics of eight non-machine learning time series forecasting methods. The relevant results are summarized in the following table:

[0047] Selection Criteria for Time Series Forecasting Methods

[0048]

[0049] Clearly, the KF method and the dual-seasonal Holt-winters method are the only two non-machine learning time series forecasting methods that can simulate trends and seasons while ensuring relatively high feasibility. Considering that the former's modeling and parameter adjustment process is very complex and far less convenient, intuitive and simple to code than the latter, and that the situation described in this invention does not require the introduction of a second seasonal factor for iteration, this invention decides to use the Holt-winters method for demand forecasting.

[0050] 2. The principle of the Holt-Winters method

[0051] The Holt-Winters method has two types: an additive model and a multiplicative model. Their mathematical expressions can be found in the relevant literature.

[0052] The additive seasonality model is suitable for situations where seasonal effects are independent of time-series level values, i.e., the amplitude of seasonal fluctuations remains constant.

[0053]

[0054] The multiplicative seasonality model is suitable for situations where the seasonal effect is proportional to the level value of the time series, meaning that the magnitude of seasonal fluctuations changes with the data level.

[0055]

[0056] In the formula: Y t This is the actual observed value at time t;

[0057] The horizontal term represents the smoothed horizontal value at time t;

[0058] T t This is the trend term, representing the trend value at time t.

[0059] S t This is the seasonal term, representing the seasonal term at time t, with a period of L;

[0060] m is a hyperparameter representing the prediction time length;

[0061] α is a hyperparameter, the smoothing coefficient of the level term, which controls the sensitivity to the latest observation;

[0062] β is a hyperparameter, the smoothing coefficient of the trend term, which controls the sensitivity to changes in the trend;

[0063] γ is a hyperparameter, the smoothing coefficient of the seasonal term, which controls the sensitivity to seasonal variations.

[0064] Since there is no clear research evidence to support the claim that demand forecasting is more suitable for additive or multiplicative models, this invention discusses both models. Regarding error assessment, the Mean Absolute Percentage Error (MAPE) is primarily used as the main indicator.

[0065]

[0066] In the formula: MAPE is a measure of relative error as a percentage;

[0067] Z i This represents the actual value for the i-th period;

[0068] This is the predicted value for period i.

[0069] Example

[0070] Figure 1 A flowchart for demand forecasting of thermal power plants based on Holt-Winters, provided for an embodiment of the present invention, specifically includes the following steps:

[0071] S1. Obtain the sales settlement data matrix within the preset historical period. The rows in the matrix represent the year, and the columns represent the month.

[0072] For example, this invention takes the prediction of the demand of Company X's thermal power plant as an example. The metadata is set as a set of sales settlement data matrix of 5 years × 12 months, where rows represent one year and columns represent one month.

[0073] S2. Input the historical data in the sales settlement data matrix into the Holt-Winters addition model and multiplication model respectively.

[0074] The sales settlement data includes the demand in the electricity market and the heat market, wherein the demand in the electricity market consists of the annual transaction volume, the market transaction volume before the month, and the market transaction volume within the month.

[0075] For special products like electric heating, supply and demand are always in balance. This is reflected in the market and in the clearing of goods. In business operations, production volume must equal demand volume and settlement volume.

[0076] For the heat market, the sales method is very simple: the monthly demand is the monthly sales volume, and it can be predicted directly using the company's historical settlement data.

[0077] For the electricity market, due to its complex sales methods, the demand in month i is always equal to the proportion of annual transactions in month i, plus the market transaction volume before and during month i.

[0078]

[0079] In the formula: These represent the total, annual, pre-month, and intra-month electricity demand for the i-th month, respectively.

[0080] However, in actual business operations, generator units inevitably experience shutdowns due to various circumstances, leading to deviations in electricity and heat demand data, thus requiring noise reduction processing. The methods for correcting these demand data also differ.

[0081] For planned production shutdowns, the criteria typically include maintenance and rescheduling. The impact of such shutdowns on sales is predictable, meaning the company knew about the need for shutdowns before formulating its sales strategy. Since the shutdown period is relatively short, current data cannot be used for recovery; instead, a moving average calculation using data from previous and subsequent periods should be employed.

[0082]

[0083] In the formula: This is for the correction requirements in month i.

[0084] For unplanned shutdowns, the criteria typically include: non-shutdown (Category I fault), load limiting (Category II fault), and RB1 (RB: RunBack), which refers to a rapid reduction in unit output due to certain reasons, known as "load shedding." The difference between RB1 and load limiting is that RB1 can generally be restored quickly. The impact of such shutdowns on sales is unpredictable; that is, the company is unaware of the need for a shutdown before formulating its sales strategy. Therefore, simple interpolation can be performed using current data.

[0085]

[0086] In the formula: This represents the number of normal working days for the unit in month i, calculated in calendar days per month.

[0087] This represents the number of abnormal days for the unit in month i.

[0088] Furthermore, a heat map of unit shutdowns was created. The method involved retrieving and tabulating the months in which units were shut down during the current year, with green squares representing normal production, yellow squares representing planned shutdowns, and red squares representing unplanned shutdowns. The numbers within each square represented the number of days of abnormal shutdown. See the statistical table of unit shutdowns for Company X from 2020-2024. Figure 2 .

[0089] This embodiment addresses the data anomaly interference caused by unit shutdown by constructing a classification and repair mechanism. It employs both moving average and linear interpolation methods to repair demand, thereby obtaining a relatively accurate set of electricity and heat demand data, providing a high-quality data foundation for global planning.

[0090] S3. Iterate through all historical data in the sales settlement data matrix, continuously fit the values ​​of the three hyperparameters in the Holt-Winters model (level, trend, and seasonality), and determine the optimal parameter combination based on the hyperparameters corresponding to the minimum mean absolute percentage error during model fitting.

[0091] First, regarding the initial values, considering the characteristics of the model, we choose data from the first two periods, such as data from January to December of the previous two years.

[0092] Regarding parameters, the level, trend, and seasonality can all be calculated iteratively by the model without requiring manual setting of values, hence the Holt-Winters model.

[0093] Regarding hyperparameters, the period length is necessarily 12, because both the model's metadata and prediction targets are measured in months; the three smoothing coefficients α, β, and γ all range from (0,1). Considering that the prediction results are crucial to the effectiveness of subsequent models, this invention employs a hierarchical calculation method during simulation, specifically including the following steps:

[0094] The first stage involved a low-precision search, where the three hyperparameters were in the range of (0,1), and a search step size of 0.1 was used to calculate a total of 113 combinations. The minimum MAPE value was then determined to be the current best combination.

[0095] The second stage involved a medium-precision stage, where the three hyperparameters were calculated in the range of ±0.1 for each parameter in the optimal combination obtained from the previous stage, with a search step size of 0.01, for a total of 213 combinations. The minimum MAPE value was then determined to be the current optimal combination.

[0096] The third stage involved a high-precision stage, which calculated a total of 213 combinations within the range of ±0.01 for the optimal combination, with a step size of 0.001. The minimum MAPE value was then determined to be the current optimal combination.

[0097] This embodiment uses an innovative "three-level progressive hyperparameter optimization strategy" to avoid the subjectivity of manual parameter tuning and ensure the model's adaptability to different seasons and market environments.

[0098] S4. Based on the optimal combination, predict the electricity and heat demand of the target thermal power plant in the next preset months.

[0099] The optimal parameter combination derived from simulations can be used to predict demand data for the next quarter. Substituting the optimal parameter combination into the model allows for the calculation of the predicted values.

[0100] If necessary, electricity demand data and heat demand data can be fitted into a virtual steam market demand data using thermodynamic principles. This makes it easier for companies to schedule production. The conversion method is as follows:

[0101]

[0102] C H,S ~C H h H

[0103] In the formula: C E C H For a certain electrical / thermal related parameter;

[0104] C E,S C H,S The steam equivalent value is derived from electrical / thermal parameters;

[0105] d0 is the power generation steam consumption rate (t·h / MW); h H Specific enthalpy of heating steam (kJ / kg).

[0106] S5. Visualized Forecast Output: Generates comparison charts of raw data, historical fitted values, and forecasts for the next 12 months. It supports intuitive comparison of results from different models, helps management quickly identify forecast trends, and improves the efficiency of monthly production scheduling by 40%.

[0107] Calculation verification:

[0108] Taking Company X as an example, the calculations of this invention can be performed using Company X's corporate data from 2020 to 2024.

[0109] The predictions were performed on a computer running a 64-bit Microsoft Windows 11 Home Chinese Edition operating system with a 12th Gen Intel(R) Core(TM) i7-12700K 3.61GHz processor and MATLAB R2023b as the model running environment.

[0110] This invention involves a total of 2 × (113 + 213 + 213) = 39706 calculations (addition + multiplication). The solution time for this model is 50.88 seconds, excluding compilation and parameter setting time. The optimal parameter combinations and their error values ​​for the two models (electricity and heat) are shown in the table below:

[0111]

[0112] The evaluation of the model's practicality revolves around MAPE (Magnetic Performance Equivalent). Literature indicates that a MAPE below 8% is considered excellent, and 10%-15% is within a reasonable range. Therefore, the MAPE value predicted in this study demonstrates that the model fits the requirements, and the results show that the exhaustive method is effective in confirming the optimal parameter selection for the prediction problem.

[0113] The simulated and fitted curves of electricity and heat demand are as follows: Figure 3 and Figure 4 Furthermore, the demand curve for steam can be constructed; see [link to relevant documentation]. Figure 5 .

[0114] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A demand forecasting method for thermal power plants based on Holt-Winters, characterized in that, include: S1. Obtain the sales settlement data matrix within a preset historical period. Rows in the matrix represent years, and columns represent months. S2. Input the historical data in the sales settlement data matrix into the Holt-Winters addition model and multiplication model respectively; S3. Iterate through all historical data in the sales settlement data matrix and continuously fit the values ​​of the three hyperparameters in the Holt-Winters model: the horizontal term, the trend term, and the seasonal term. Determine the optimal parameter combination based on the hyperparameters corresponding to the minimum mean absolute percentage error during the model fitting process. S4. Based on the optimal combination, predict the electricity and heat demand of the target thermal power plant in the next preset months.

2. The method according to claim 1, characterized in that, S3 includes: When the three hyperparameters are in the range (0, 1), with the first preset step size of 0.1 as the search step size, the minimum mean absolute percentage error value obtained by the search is taken as the first optimal combination parameter. Centered on the hyperparameters corresponding to the first optimal combination, a traversal search is performed within the range of ±0.1 with a second predetermined step size of 0.01, and the minimum mean absolute percentage error value obtained from the search is taken as the second optimal combination parameter. Centered on the hyperparameters corresponding to the second optimal combination, a traversal search is performed within the range of ±0.01 with a third predetermined step size of 0.001, and the minimum mean absolute percentage error value obtained from the search is taken as the final optimal parameter combination.

3. The method according to claim 1, characterized in that, The sales settlement data includes the demand in the electricity market and the heat market, wherein the demand in the electricity market consists of the annual transaction volume, the market transaction volume before the month, and the market transaction volume within the month.

4. The method according to claim 3, characterized in that, For electricity market demand, when planned shutdowns occur, the monthly demand is adjusted using a moving average of previous and subsequent data. The adjustment formula is as follows: In the formula, This is for the correction requirements in month i.

5. The method according to claim 3, characterized in that, For electricity market demand, when unplanned shutdowns occur, an interpolation correction is applied based on the ratio of normal working days to abnormal working days. The correction formula is as follows: In the formula, This represents the number of normal working days for the unit in month i, calculated in calendar days per month. This represents the number of abnormal days for the unit in month i.

6. The method according to claim 1, characterized in that, Thermodynamic principles are used to fit electricity demand data and heat demand data into a virtual steam market demand data set. The fitting method is as follows: C H,S ~C H h H In the formula, C E Let C be the demand parameter for a given electricity market. H For the demand parameters of a certain heat market; C E,S The steam equivalent value converted from the demand parameters of the electricity market; C H,S The steam equivalent value converted from the demand parameters of the heat market; d0 is the power generation steam consumption rate; h H Specific enthalpy of heating steam.

7. The method according to claim 1, characterized in that, Following S4, the method further includes: Plot a comparison graph of the original data, historical fitted values, and future predicted values, and use different colors and line styles to distinguish the results of the additive model and the multiplicative model.

Images