AI-based attendance forecasting

This post is part of a larger article: AI in the service of heritage: Droplabs Dynamic Pricing and attendance prediction in shaping sustainable tourist behavior in Tatra National Park.

‍

At first glance, predicting attendance at the Palenica Białczańska parking lot seems ridiculously simple. A perfectly sunny weekend or a public holiday? You can bet without looking that the parking lot will be bursting at the seams! An all-day downpour and fog on an autumn weekday? Exactly - it will be empty. Predicting occupancy on those dozen or so extreme, “black and white” days of the year is no great art when extreme weather and obvious calendar factors coincide.

‍

The art of prediction, however, does not lie in determining occupancy on these few, easily identifiable days. The real challenge begins with the “non-obvious” days that make up most of the calendar: how do you predict occupancy when only partial cloud cover is forecast for the weekend? What determines attendance on a summer holiday day when total rainfall is just above zero, or when the maximum temperature drops below 20°C? On top of that come ordinary weekdays outside the holidays, which make up the largest share of the year.

‍

To better illustrate this challenge, let's analyze a scatter plot showing attendance at the Palenica Białczańska parking lot in relation to total daily rainfall. The parking lot's maximum capacity is about 1,300 cars.

‍

‍

Each dot represents one day of the year, and their distribution immediately refutes simple intuitions. On days with no precipitation, attendance ranged from close to zero up to the maximum. These days are the dots on the vertical line at 0 mm of total daily precipitation. Likewise, there are many days with up to 10 mm of rainfall on which attendance was very high. Large variations in attendance can also be seen on days with even more rainfall.

‍

“Pearson's coefficient measures linear correlation, whose hypothetical course is plotted on the chart as a black line. This does not mean that customers' actual behavior is linear; with this method, we disregard other types of correlations and dependencies. With such a distribution of points, the Pearson coefficient indicates a weak relationship. We are dealing with a random distribution with a slight downward trend. This feature can serve as additional information for fine-tuning the predictive model, but on its own, without the context of other features, it does not allow us to draw any conclusions.”

^{- says Michał Piłat, Data Scientist at Droplabs}

‍

Similar examples abound when analyzing the data by other features, such as snow cover or cloud cover, or when looking only at selected periods, such as school holidays, winter breaks in specific regions, or public holidays.

‍

“Humans have an innate tendency to remember distinctive events and give them more weight than statistics would justify. Unlike human perception, the machine learning (AI) systems developed by Droplabs and used in Tatra National Park make it possible to analyze real, extensive data sets in multidimensional contexts. This allows them to identify subtle dependencies and patterns that go beyond schematic thinking and intuitive simplifications.”

^{- adds Monika Karpierz-Nałęcz, Head of AI at Droplabs}

‍

To meet the challenge of predicting attendance, Tatra National Park uses the AI-based Attendance Forecast provided by Droplabs. The web panel shows the expected attendance for a given future day as a “from-to” range. To help users understand how accurate the forecast is, the panel also shows past dates, where the historical forecast can be compared with the actual values. This helps build confidence in the models.

‍

‍

Like standard weather forecasts, Droplabs attendance predictions for future days are not static; they change over time. The rule is simple: the closer the date, the more accurate the prediction. This is because the time horizon gets shorter and the models have more and better-quality data, such as fresher, more accurate weather forecasts and more signals about customer behavior in the online store.

‍

“Longer prediction horizons are inherently harder to estimate, because the model has to process and verify the entire, dynamically changing data context. For example, the same booking level can mean different things depending on the weather forecast, suggesting either planned or spontaneous visits. When the weather forecast changes, the entire decision-making context changes. The model needs to match its prediction to the current market situation and external conditions, rather than rely on data and assumptions from a few days earlier.”

^{- points out Piotr Pyznarski, Droplabs founder responsible for product and technology (CTO)}

‍

Droplabs predictive models are highly accurate. In Tatra National Park, they correctly determine the attendance range for the next day in about 90% of cases.

‍

‍

As an illustration, let's look at the forecast history for the second half of September 2025 shown above. For Friday, September 19, 2025, the following weekend, and the working Monday, September 22, the model forecast a full parking lot, and that is exactly what happened.

‍

More importantly, the system successfully captures subtle changes in demand. The model correctly predicted a significant drop in attendance on Thursday, September 25, compared with the previous Thursday. The actual data confirmed this trend: only 350 cars were recorded on September 25, while the Thursday a week earlier attracted more than 800 vehicles.

‍

‍

The data from October 9, 2025 shows that demand does not depend on the weather forecast alone. On Thursday, October 9, in rainy and cool weather, 217 cars arrived at the parking lot. Even though practically identical weather was forecast for the following Friday, the attendance model predicted a much higher range. The actual data confirmed this prediction: attendance on Friday more than doubled, reaching 450 vehicles.

‍

The TPN parking facilities on the Morskie Oko trail are a complex infrastructure made up of three separate parking zones: the main parking lot at Palenica Białczańska (with a capacity of about 600 vehicles), the roadside lane between Palenica Białczańska and Łysa Polana, and a paved parking lot at Łysa Polana.

‍

Accurate attendance forecasts are important for TPN's operational preparations. They enable efficient planning of Park Service patrols, optimal deployment of staff at the parking lots, and ongoing adjustment of communication with tourists. As a result, more efficient operations management helps optimize the Park's running costs.

‍

In the next article, we take a closer look at the technology behind these solutions and at Droplabs AI systems. You can read it here.

‍

This post is part of a larger article: AI in the service of heritage: Droplabs Dynamic Pricing and attendance prediction in shaping sustainable tourist behavior in Tatra National Park.

‍