Time Series Forecasting for Touristic Policies ^†

Abstract

The formulation of touristic policies is a time-consuming process that consists of a wide range of steps and procedures. These policies are highly dependent on the number of tourists and visitors to an area to be as effective as possible. The estimation of this number is not always easy to achieve, since there is a lack of the corresponding data (i.e., number of visitors per day). Hence, this estimation must be achieved by utilizing alternative data sources. To this end, in this paper, the authors propose a neural network architecture that is trained on waste management data to estimate the number of visitors and tourists in the highly touristic municipality of Vari-Voula-Vouliagmeni, Greece.

1. Introduction

According to [1], Greece is consistently in the top ten (10) countries in the European Union with the highest number of international tourist arrivals. There exists a plethora of policies at national and regional level that aim to manage the increasing number of visitors and tourists in the most efficient way possible to (i) ensure the highest quality of the services provided and (ii) ease the everyday lives of the locals during touristic seasons. However, the development and evaluation of those policies is an ongoing and time-consuming process that highly depends on provided historical data with regard to arrivals and overnight stays.

In Greece, at the national level, the abovementioned data are available from the Hellenic Statistical Authority (e.g., arrivals in the Athens International Airport) [2]. At the local level (e.g., at the level of a municipality), these data are mostly collected by private businesses, such as hotels, so access to them is granted after complex bureaucratic procedures. As a result, the formulation of touristic policies at a local level is complicated and requires additional effort and resources to collect the appropriate data. However, in the last couple of years, several municipalities in Greece have introduced smart sensors in their operations, thus accumulating data for digitalizing and/or improving their processes. Those sensor data cover a variety of domains, such as waste management, which, even though they do not directly relate to the tourism domain, they can, however, provide insights with regard to visitors and tourists’ flows.

An indicative example of such a Greek municipality, is the municipality of Vari-Voula-Vouliagmeni, or VVV for short. VVV has a waste management process in place wherein smart bins have been installed in the majority of the municipality, such as outside hotels and restaurants [3]. Those smart bins contain sensors that take hourly measurements of the volume of waste that is present inside. Policy makers in the municipality of VVV plan to be use such data to facilitate the process of garbage collection by providing optimized routes for the corresponding garbage trucks [4]. However, the fill level of the bins can also be utilized in the context of touristic policies, since the rate at which a bin fills is highly correlated with the flow of citizens/tourists in the area where this bin is located. In other words, the smart bins that are located in more populated areas and areas that are visited by more people fill at a higher rate than the smart bins that are located in less populated areas.

To this end, in this paper, the authors propose a machine learning approach for predicting the touristic flows in VVV. More specifically, firstly, a neural network is trained to provide predictions for the future fill level of the smart bins. Then, the K-means clustering algorithm is utilized to (i) split the municipality into a number of areas based on the location of the bins and (ii) cluster those areas based on the predicted fill levels and the corresponding fill rate. Then, the proposed approach generates a map that highlights, with different colors, the clustered areas in terms of the projected fill rate, thus allowing the policy makers to identify the expected number of tourists per area of interest and create/enhance the corresponding policies.

The rest of this paper is organized as follows: Section 1 is the Introduction, where the initial problem statement is provided and the proposed approach is briefly described. Section 2 provides a state-of-the-art analysis with regard to time series forecasting and machine learning in the context of smart cities. Section 3 provides a detailed description of the dataset and techniques used, as well as an overview of the proposed approach. In Section 4, the results of the proposed approach are presented, and they are discussed further in Section 5. Lastly, Section 6 summarizes this manuscript and provides further insights with respect to future steps.

2. Literature Review

2.1. Time Series Forecasting

In Computer Science, time series forecasting refers to the analysis of data that are time-dependent and the application of techniques and algorithms of diverse complexity in order to predict specific steps related to the future based on a set of past values of the said data [5]. There exist two (2) types of approaches for performing time series forecasting. The first one is based on statistical methods, and the second one utilizes more complex machine learning (i.e., ML) algorithms, such as neural networks. In general, statistical methods and less complex ML algorithms are suitable for datasets that do not contain complex temporal dependencies, whilst more complex ML models are capable of capturing those dependencies [6]. With regard to statistical methods, the most common are autoregressive integrated moving average (i.e., ARIMA), ARIMA with eXogenous factors (i.e., ARIMAX), ARIMAX with seasonal terms (i.e., SARIMAX), antileakage least-squares spectral analysis (i.e., ALLSSA) [7], Naive2, Theta, exponential smoothing (i.e., ETS), and Trigonometric Exponential smoothing state space model with Box–Cox transformation, ARMA errors, Trend and Seasonal components (i.e., TBATS). Traditional ML algorithms that have been used in the literature to solve time series forecasting problems include K-nearest neighbor (i.e., KNN) [8] and support vector machine (i.e., SVM) models [9].

As for more complex ML models that have been introduced in solving tasks related to time series forecasting, those mainly consist of different types of neural network architectures, including convolutional neural networks (i.e., CNNs), recurrent neural networks (i.e., RNNs), and transformers. CNNs are usually utilized in computer vision-related use cases, but they have also been used for time series forecasting in specific domains, such as finance, showcasing promising results and even outperforming other traditional approaches [10]. RNNs, such as long short-term memory (i.e., LSTM) models and gated recurrent units (i.e., GRUs), are ideal for time series forecasting since they are able to capture the temporal dependencies present in the data, as they contain an implicit memory of previous inputs [11]. Moreover, RNNs are often combined with statistical methods such as ARIMA, thus outperforming non-hybrid approaches [12]. Lastly, transformers, which have achieved unparallel performance in natural language processing (i.e., NLP) tasks, have also been recently adopted in solving time series forecasting problems since they are able to capture long-range dependencies [13].

2.2. Time Series Forecasting in Smart Cities

The increasing number of sensors that are being installed in cities and municipalities has resulted in the generation of a vast amount of times series data. Those data can be utilized to train ML models that are able to provide forecasts, thus allowing public workers and policy makers to take data-driven decisions. There exist numerous examples in the literature of approaches that perform time series forecasting to support decision making in different aspects of a smart city, such as waste management and water management.

More specifically, as regards waste management, the authors in [14] carried out a comparative study with regard to five (5) different applications of ML in this domain, namely municipal solid waste management; composting; anaerobic digestion; incineration, pyrolysis, and gasification; and landfill. They concluded that, in the literature, the most frequently employed algorithms are CNNs and RNNs, followed by SVMs, the genetic algorithm (i.e., GA), decision trees (i.e., DTs), and the random forest (i.e., RF) algorithm. Indeed, the authors in [15], make use of an LSTM model to perform municipal solid waste (i.e., MSW) multi-classification, achieving high accuracy in terms of the root mean squared error (i.e., RMSE), whilst the authors in [16] also prefer another type of an artificial neural network (i.e., ANN) for forecasting MSW, because NNs and specifically LSTMs are preferred for MSW forecasting over more traditional approaches, since they provide the highest accuracy [17].

With regard to water management, there exist numerous approaches in the literature that mainly focus on drinking water and wastewater management. As for the latter, the authors in [18] experiment on a set of different NN architectures. According to their experiments and corresponding results, the most appropriate architectures for forecasting the energy consumption of a wastewater treatment plant (i.e., WWTP) were found to be the linear and dense NNs, since the complexity of the CNN and RNN did not improve the accuracy of the forecasts. Similar approaches that utilize times series data from WWTPs to forecast parameters related to their performance are also presented in [19], where the authors also compared three (3) different NN architectures, and in [20], where the authors compared CNNs and RNNs and determine that the CNNs achieved the highest accuracy on their data.

As for tourism, there exists several publications, such as [21,22,23], which utilize several methods to forecast the arrival of tourists in an area. In those approaches, the time series data that are being utilized originate from sources that directly count the number of tourists/visitors, such as airports. However, providing forecasts with regard to the future number of tourists that will visit an area based on indirect data that are potentially good indicators of this number should be considered. That way, the developed touristic policies could be further improved, since such data could provide a more detailed view of tourist flow in a more specific area. To this end, in this paper, data regarding waste management are utilized, and the appropriate ML methods are applied to estimate tourist flow in the highly touristic Greek municipality of Vari-Voula-Vouliagmeni.

3. Materials and Methods

3.1. Dataset

As previously mentioned, the data used for this approach originate from smart bins located in the VVV municipality, since the municipality does not own any data related to daily arrivals of tourists. The presented approach is based on the fact that the volume of waste that a bin contains is highly correlated with the population density of the corresponding area. This means that the bins of a highly touristic area fill faster than the ones located in a less touristic one.

More specifically, the said smart bins contain sensors that measure the volume of the waste that they contain. The majority of bins measure the volume every hour, whilst five (5) of them do so every ten (10) minutes. The total number of bins that are currently installed in the municipality is one thousand (1000). A visualization of the bins on a map of the municipality is provided in Figure 1, with the bins highlighted in blue.

The dataset consists of the timestamp column, which represents the date when the sensor took the corresponding measurement; the latitude and longitude columns, representing the location of every bin; and the volume column, which refers to the related volume of waste that a bin contains, measured in liters. The date range of the dataset is approximately two (2) years, starting from August 2022 up to October 2024.

3.2. Preprocessing and Data Cleaning

Prior to training the corresponding time series forecasting model, specific preprocessing and data cleaning steps had to take place in order to ensure the reliability of the dataset and minimize the data bias that could be introduced into the ML model. Firstly, since some smart bins measure the volume of the waste every ten (10) minutes whilst the majority of the bins measure it every hour, some bins contained six (6) observations per hour. As a result, the latter were sampled to only keep the observation in line with the maximum value per hour, since their number was negligible compared to the total number of bins, and thus, their down sampling did not affect the accuracy of the predictive model. The selection of the maximum value is considered to handle the event of a garbage truck emptying the bin and provide a better idea of the rate at which a bin fills. Moreover, a technique called Fourier smoothing was applied on the volume column to smooth the corresponding time series and remove any potential “noise” [24], whilst capturing the underlying seasonal variations in the data.

As soon as the preprocessing step was complete, the data cleaning process took place to ensure the reliability of the data and deal with any potential inconsistencies present in the dataset. The cleaning process that was adopted is a variation of the one that was presented in [25,26]. In detail, the way that this cleaning process functions is that a data schema is generated for the dataset in question, and then the dataset is validated based on this schema. In general terms, the quality of the dataset was sufficient, with no major issues. However, in the dataset, there existed a few missing values, which were replaced based on the previous and next observation. More specifically, each missing value was replaced with the mean value of the previous and next non-missing values. The number of missing values was insignificant; thus, no notable change in the data variance was identified.

3.3. Algorithms

Given the fact that RNNs are widely used in the literature for time series forecasting, two (2) different types of the said NNs were implemented and compared in the context of this manuscript. In deeper detail, the first architecture is an LSTM model, and the second one is a GRU. Both architectures used forty-eight (48) records as inputs to provide forecasts for the next twenty-four (24) hours. Early stopping was applied during the training to avoid overfitting, while several optimizers were tested, since it is known that the selection of the optimizer directly affects an NN in terms of bias and accuracy [27,28]. The optimizers that were utilized in the aforementioned architectures included Adam, Stochastic Gradient Descend (i.e., SGD), AdamW, Adamax, Adagrad, Adadelta, Adafactor, RMSprop, and Nadam, which are available through the python Darts module specifically designed for time series forecasting tasks [29]. The metric that was calculated to assess the performance of the different architectures was the mean absolute error (i.e., MAE).

The K-means clustering algorithm was used to (i) split the municipality into a number of areas, based on the location of the bins, and (ii) cluster those areas based on the predicted fill levels and the corresponding fill rate. K-means is one of the simplest and fastest clustering algorithms that exists [30,31], which is why it was selected for the proposed approach. Moreover, the experimental results that were generated were reviewed by experts who are familiar with the municipality, thus confirming that the output of the algorithm, in terms of clustered areas, is valid.

3.4. Proposed Approach Architecture

Overall, the proposed approach follows a specific flow, as shown in Figure 2.

More specifically, the data are collected with the help of appropriate application programming interfaces (APIs). Then, specific preprocessing and data cleaning tasks take place to ensure the reliability of the data, thus dealing with any potential data bias, as described in Section 3.2. Afterwards, the NN is trained on these data to provide predictions with regard to the future fill levels of the bins. It is also worth noting that the model updates itself whenever new data are available so that it is able to recognize the latest trends in the dataset. When the predictions for every bin are provided by the model, the rate at which each bin will fill is calculated based on them. Given the fact that the predicted fill rate for every bin is available, the K-means algorithm is applied. The optimal number of clusters/regions in which K-means will split the municipality is determined by the elbow method. Furthermore, the number of clusters in which those regions will be grouped, based on the predicted fill levels and the corresponding fill rate, is also determined by the elbow method. However, the users can also use a number of clusters of their choice. Finally, an interactive map is generated, where the different areas and bins groups are highlighted, providing a list in descending order with regard to the corresponding fill rate of every group of bins. Since the fill rate of the bins is a strong indicator of the touristic flows, this visualization showcases which areas are more and less touristic, thus allowing policy makers to make better, data-driven decisions. An indicative example of such a visualization is depicted in Figure 3a, where the bins in areas “10” and “3” are projected to have the greatest number of tourists, followed by area “1”. Moreover, an indicative visualization of the time series for two (2) bins for a small period is shown in Figure 3b, where the forecasted values are marked in red and the actual values per bin are shown in blue.

4. Results

With regard to the selection of the most appropriate NN architecture for predicting the future fill levels of the bins, which is the key aspect of the proposed approach, a set of experiments were carried out to identify the most appropriate NN architecture. The results are summarized in

Table 1. Experimental results.

NN Type	Optimizer	Learning Rate	Epochs	MAE	Training Time
LSTM	Adam	0.01	3	0.0052	3.03
		0.001	3	0.00489	4.09
		0.0001	3	0.014	3.57
	SGD	0.01	3	0.0283	2.57
		0.001	3	0.196	3.30
		0.0001	5	0.253	4.55
	AdamW	0.01	3	0.00363	3.15
		0.001	3	0.00542	3.15
		0.0001	3	0.00861	3.18
	Adamax	0.01	3	0.00324	2.57
		0.001	3	0.00622	3.24
		0.0001	3	0.0117	3.09
	Adagrad	0.01	3	0.00943	3.12
		0.001	3	0.0467	2.57
		0.0001	12	0.274	13.10
	Adadelta	0.01	3	0.0368	2.57
		0.001	5	0.206	4.50
		0.0001	16	0.313	18.24
	Adafactor	0.01	3	0.00902	3.39
		0.001	3	0.00568	3.03
		0.0001	4	0.0668	4.36
	RMSprop	0.01	3	0.00928	3.03
		0.001	3	0.00396	3.03
		0.0001	3	0.0056	3.09
	Nadam	0.01	3	0.00401	2.57
		0.001	3	0.00439	3.03
		0.0001	3	0.0148	3.12
GRU	Adam	0.01	3	0.0032	3.03
		0.001	3	0.00418	2.44
		0.0001	3	0.0112	2.57
	SGD	0.01	3	0.0238	2.54
		0.001	3	0.0469	2.51
		0.0001	5	0.216	4.45
	AdamW	0.01	3	0.0053	2.51
		0.001	3	0.00411	2.54
		0.0001	3	0.0111	2.51
	Adamax	0.01	3	0.00318	2.51
		0.001	3	0.00523	2.51
		0.0001	3	0.00972	2.51
	Adagrad	0.01	3	0.00599	3.12
		0.001	3	0.0359	3.00
		0.0001	9	0.326	9.45
	Adadelta	0.01	3	0.0302	2.46
		0.001	4	0.187	3.44
		0.0001	10	0.337	10.10
	Adafactor	0.01	3	0.0135	2.51
		0.001	3	0.0067	2.48
		0.0001	5	0.0351	4.20
	RMSprop	0.01	3	0.0144	2.51
		0.001	3	0.00645	3.03
		0.0001	3	0.00629	2.45
	Nadam	0.01	3	0.0068	3.09
		0.001	3	0.00521	3.21
		0.0001	3	0.00919	2.57

. More specifically, the following table showcases the calculated MAE; the training time needed (in minutes); and the number of epochs needed per NN architecture, optimizer, and learning rate.

5. Discussion

According to the results of the experiments, with regard to the LSTM NN, the most appropriate combination of optimizer and learning rate in terms of MAE is the Adamax optimizer with a learning rate of 0.01, which achieved an MAE of 0.00324. As for the GRU NN, the most appropriate combination of optimizer and learning rate in terms of MAE is the Adamax optimizer with a learning rate of 0.01, which achieved an MAE of 0.00318. Overall, the experiments showed that, for this specific use case, the best architecture is the GRU NN that utilizes the Adamax optimizer with a learning rate of 0.01, since it not only performed best in terms of accuracy but also generally outperformed LSTM in terms of training speed [32]. As for the clustering algorithm, K-means provided sufficient groups that were validated by corresponding experts, thus succeeding in both (i) splitting the municipality in several areas, based on the location of the bins, and (ii) clustering those areas based on the predicted fill levels and the corresponding fill rate. However, it should be mentioned that the proposed approach would benefit from experimentation with other clustering algorithms, as well as smoothing techniques, to assess whether they can assist the trained model in providing even more accurate predictions.

6. Conclusions

In conclusion, predicting the touristic flows in an area is of vital importance for the formulation of the corresponding policies that will enhance both the experiences of the visitors and ease the everyday lives of the locals. However, the data needed for such predictions (e.g., number of visitors per day) are not always available. As a result, alternative data sources must be utilized. Based on this, this paper showcased how waste management data and, more specifically, the fill levels of the smart bins of a municipality can be utilized to estimate touristic flows. A specific ML pipeline was implemented, and several experiments were carried out. Those experiments identified that the combination comprising a GRU NN utilizing the Adamax optimizer with a learning rate of 0.01 and the K-means algorithm is able to accurately provide forecasts and cluster the municipality into areas of interest for policy makers.

As regards future steps, the authors aim to further experiment with other aspects of the proposed approach. More specifically, the authors plan to utilize other time series smoothing techniques. Moreover, even though K-means provided adequate results, they plan to test other clustering algorithms and evaluate their performance. Lastly, it would be beneficial to utilize the proposed approach in other similar use cases, potentially with datasets from other municipalities, to further validate the generated results.