|
Namangan Institute of Engineering and Technology nammti uz
Pdf ko'rish
|
| bet | 313/693 | | Sana | 13.05.2024 | | Hajmi | 15,56 Mb. | | #228860 |
Bog'liq ТўпламMethods. We developed a machine learning (ML) model that predicts core indicators of
frequency stability using techno-economic factors in order to model frequency stability and control
based on operational data. Time series data, including load, generation, and electricity price time
series, were utilized as inputs [9]. The time series includes day-ahead accessible data, including
generation projections, which may be used to predict frequency stability one day in advance, as well
as ex-post (actually observed) data. We created stability indicators using grid frequency
measurements [9, 10, 11] in the synchronous regions of Great Britain, Continental Europe, and the
Nordic as outputs (targets). Instead, we utilized the active control volumes as output data to
simulate control activation in Germany [12]. All data sources are openly accessible, and the data has
an hourly or 15-minute resolution. Using a Gradient Boosted Tree (GBT) model to fit the data, we
were able to map inputs to outputs [5]. Next, we extracted significant input characteristics and
dependencies from the machine learning model and explained it using Shapely Additive
Explanations (SHAP), a technique from XAI [9]. This makes transparent day-ahead forecasts possible
and allows for extensive ex-post interpretations of historical operational data.
Results. Our approach identified the primary European power systems' frequency stability
drivers and hazards [12]. Fast generation ramps primarily affect frequency gradients, and we
identified three distinct kinds of generation types with opposing effects. Electricity pricing and
prediction mistakes are the main factors influencing the stability indicator, which gauges the
necessary management effort. Forecast mistakes often account for a large portion of frequency
variances in the Nordic grid, quick load and generation ramps dominate the Continental European
system, and renewable energy and electricity pricing play a significant influence in Great Britain.
Additionally, our method clarified how various load and generation parameters lead to
deterministic frequency deviations (DFDs) [6]. A significant portion of the daily DFD pattern is
already described by the daily cycle of the load, namely the load gradients. But this strategy doesn't
work well, particularly before and after midday. We linked this observation to solar power ramps
by utilising a variety of input characteristics, and we created a model that almost exactly replicates
the daily DFD pattern in the data.
Lastly, we examined the historical activation of control power in Germany and its forecast for
the next day using our methodology [11]. We demonstrated that although prediction errors and the
generation mix have a significant impact on the control power's activation, it only displays modest
daily trends. Sufficient prediction error estimates seem to be a crucial element in enhancing the
control's day-ahead predictability. Generally speaking, either ex-post analysis or day-ahead
prediction requires the employment of different loss functions and model training techniques.
|
| |
